diff --git a/.devops/cloud-v-pipeline b/.devops/cloud-v-pipeline
deleted file mode 100644
index af8c0cea6155c..0000000000000
--- a/.devops/cloud-v-pipeline
+++ /dev/null
@@ -1,22 +0,0 @@
-node('x86_runner1'){            // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
-    stage('Cleanup'){
-        cleanWs()               // Cleaning previous CI build in workspace
-    }
-    stage('checkout repo'){
-        retry(5){               // Retry if the cloning fails due to some reason
-            checkout scm        // Clone the repo on Runner
-        }
-    }
-    stage('Compiling llama.cpp'){
-        sh'''#!/bin/bash
-            make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
-        '''
-    }
-    stage('Running llama.cpp'){
-        sh'''#!/bin/bash
-            module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
-            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
-            cat llama_log.txt                   # Printing results
-        '''
-    }
-}
diff --git a/.devops/cpu.Dockerfile b/.devops/cpu.Dockerfile
index 9459f08c10c94..e1bb7d4675dc3 100644
--- a/.devops/cpu.Dockerfile
+++ b/.devops/cpu.Dockerfile
@@ -4,8 +4,6 @@ FROM ubuntu:$UBUNTU_VERSION AS build
 
 ARG TARGETARCH
 
-ARG GGML_CPU_ARM_ARCH=armv8-a
-
 RUN apt-get update && \
     apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
@@ -13,10 +11,8 @@ WORKDIR /app
 
 COPY . .
 
-RUN if [ "$TARGETARCH" = "amd64" ]; then \
+RUN if [ "$TARGETARCH" = "amd64" ] || [ "$TARGETARCH" = "arm64" ]; then \
         cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
-    elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
     else \
         echo "Unsupported architecture"; \
         exit 1; \
diff --git a/.devops/cuda.Dockerfile b/.devops/cuda.Dockerfile
index 94f143397233f..4b708ae278ddf 100644
--- a/.devops/cuda.Dockerfile
+++ b/.devops/cuda.Dockerfile
@@ -61,7 +61,7 @@ RUN apt-get update \
     python3 \
     python3-pip \
     && pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt \
+    && pip install --break-system-packages -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
diff --git a/.devops/vulkan.Dockerfile b/.devops/vulkan.Dockerfile
index fcd81ffa1e94e..6cf87c67e8553 100644
--- a/.devops/vulkan.Dockerfile
+++ b/.devops/vulkan.Dockerfile
@@ -2,14 +2,30 @@ ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
-# Install build tools
-RUN apt update && apt install -y git build-essential cmake wget
+# Ref: https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html
 
-# Install Vulkan SDK and cURL
-RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
-    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list https://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
-    apt update -y && \
-    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget xz-utils
+
+# Install Vulkan SDK
+ARG VULKAN_VERSION=1.4.321.1
+RUN ARCH=$(uname -m) && \
+    wget -qO /tmp/vulkan-sdk.tar.xz https://sdk.lunarg.com/sdk/download/${VULKAN_VERSION}/linux/vulkan-sdk-linux-${ARCH}-${VULKAN_VERSION}.tar.xz && \
+    mkdir -p /opt/vulkan && \
+    tar -xf /tmp/vulkan-sdk.tar.xz -C /tmp --strip-components=1 && \
+    mv /tmp/${ARCH}/* /opt/vulkan/ && \
+    rm -rf /tmp/*
+
+# Install cURL and Vulkan SDK dependencies
+RUN apt install -y libcurl4-openssl-dev curl \
+    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev
+
+# Set environment variables
+ENV VULKAN_SDK=/opt/vulkan
+ENV PATH=$VULKAN_SDK/bin:$PATH
+ENV LD_LIBRARY_PATH=$VULKAN_SDK/lib:$LD_LIBRARY_PATH
+ENV CMAKE_PREFIX_PATH=$VULKAN_SDK:$CMAKE_PREFIX_PATH
+ENV PKG_CONFIG_PATH=$VULKAN_SDK/lib/pkgconfig:$PKG_CONFIG_PATH
 
 # Build it
 WORKDIR /app
diff --git a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
index 95a0b5cc75bde..feb0d512055a6 100644
--- a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
+++ b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
@@ -40,7 +40,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
         multiple: true
     validations:
       required: true
diff --git a/.github/ISSUE_TEMPLATE/011-bug-results.yml b/.github/ISSUE_TEMPLATE/011-bug-results.yml
index d1034bbb6910e..c42a14ff83eb6 100644
--- a/.github/ISSUE_TEMPLATE/011-bug-results.yml
+++ b/.github/ISSUE_TEMPLATE/011-bug-results.yml
@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
         multiple: true
     validations:
       required: true
diff --git a/.github/copilot-instructions.md b/.github/copilot-instructions.md
new file mode 100644
index 0000000000000..3250e3279ecb6
--- /dev/null
+++ b/.github/copilot-instructions.md
@@ -0,0 +1,262 @@
+# Copilot Instructions for llama.cpp
+
+## Repository Overview
+
+llama.cpp is a large-scale C/C++ project for efficient LLM (Large Language Model) inference with minimal setup and dependencies. The project enables running language models on diverse hardware with state-of-the-art performance.
+
+**Key Facts:**
+- **Primary language**: C/C++ with Python utility scripts
+- **Size**: ~200k+ lines of code across 1000+ files
+- **Architecture**: Modular design with main library (`libllama`) and 40+ executable tools/examples
+- **Core dependency**: ggml tensor library (vendored in `ggml/` directory)
+- **Backends supported**: CPU (AVX/NEON optimized), CUDA, Metal, Vulkan, SYCL, ROCm, MUSA
+- **License**: MIT
+
+## Build Instructions
+
+### Prerequisites
+- CMake 3.14+ (primary build system)
+- C++17 compatible compiler (GCC 13.3+, Clang, MSVC)
+- Optional: ccache for faster compilation
+
+### Basic Build (CPU-only)
+**ALWAYS run these commands in sequence:**
+```bash
+cmake -B build
+cmake --build build --config Release -j $(nproc)
+```
+
+**Build time**: ~10 minutes on 4-core system with ccache enabled, ~25 minutes without ccache.
+
+**Important Notes:**
+- The Makefile is deprecated - always use CMake
+- ccache is automatically detected and used if available
+- Built binaries are placed in `build/bin/`
+- Parallel builds (`-j`) significantly reduce build time
+
+### Backend-Specific Builds
+For CUDA support:
+```bash
+cmake -B build -DGGML_CUDA=ON
+cmake --build build --config Release -j $(nproc)
+```
+
+For Metal (macOS):
+```bash
+cmake -B build -DGGML_METAL=ON
+cmake --build build --config Release -j $(nproc)
+```
+
+**Important Note**: While all backends can be built as long as the correct requirements for that backend are installed, you will not be able to run them without the correct hardware. The only backend that can be run for testing and validation is the CPU backend.
+
+### Debug Builds
+Single-config generators:
+```bash
+cmake -B build -DCMAKE_BUILD_TYPE=Debug
+cmake --build build
+```
+
+Multi-config generators:
+```bash
+cmake -B build -G "Xcode"
+cmake --build build --config Debug
+```
+
+### Common Build Issues
+- **Issue**: Network tests fail in isolated environments
+  **Solution**: Expected behavior - core functionality tests will still pass
+
+## Testing
+
+### Running Tests
+```bash
+ctest --test-dir build --output-on-failure -j $(nproc)
+```
+
+**Test suite**: 38 tests covering tokenizers, grammar parsing, sampling, backends, and integration
+**Expected failures**: 2-3 tests may fail if network access is unavailable (they download models)
+**Test time**: ~30 seconds for passing tests
+
+### Server Unit Tests
+Run server-specific unit tests after building the server:
+```bash
+# Build the server first
+cmake --build build --target llama-server
+
+# Navigate to server tests and run
+cd tools/server/tests
+source ../../../.venv/bin/activate
+./tests.sh
+```
+**Server test dependencies**: The `.venv` environment includes the required dependencies for server unit tests (pytest, aiohttp, etc.). Tests can be run individually or with various options as documented in `tools/server/tests/README.md`.
+
+### Test Categories
+- Tokenizer tests: Various model tokenizers (BERT, GPT-2, LLaMA, etc.)
+- Grammar tests: GBNF parsing and validation
+- Backend tests: Core ggml operations across different backends
+- Integration tests: End-to-end workflows
+
+### Manual Testing Commands
+```bash
+# Test basic inference
+./build/bin/llama-cli --version
+
+# Test model loading (requires model file)
+./build/bin/llama-cli -m path/to/model.gguf -p "Hello" -n 10
+```
+
+## Code Quality and Linting
+
+### C++ Code Formatting
+**ALWAYS format C++ code before committing:**
+```bash
+git clang-format
+```
+
+Configuration is in `.clang-format` with these key rules:
+- 4-space indentation
+- 120 column limit
+- Braces on same line for functions
+- Pointer alignment: `void * ptr` (middle)
+- Reference alignment: `int & ref` (middle)
+
+### Python Code
+**ALWAYS activate the Python environment in `.venv` and use tools from that environment:**
+```bash
+# Activate virtual environment
+source .venv/bin/activate
+```
+
+Configuration files:
+- `.flake8`: flake8 settings (max-line-length=125, excludes examples/tools)
+- `pyrightconfig.json`: pyright type checking configuration
+
+### Pre-commit Hooks
+Run before committing:
+```bash
+pre-commit run --all-files
+```
+
+## Continuous Integration
+
+### GitHub Actions Workflows
+Key workflows that run on every PR:
+- `.github/workflows/build.yml`: Multi-platform builds
+- `.github/workflows/server.yml`: Server functionality tests
+- `.github/workflows/python-lint.yml`: Python code quality
+- `.github/workflows/python-type-check.yml`: Python type checking
+
+### Local CI Validation
+**Run full CI locally before submitting PRs:**
+```bash
+mkdir tmp
+
+# CPU-only build
+bash ./ci/run.sh ./tmp/results ./tmp/mnt
+```
+
+**CI Runtime**: 30-60 minutes depending on backend configuration
+
+### Triggering CI
+Add `ggml-ci` to commit message to trigger heavy CI workloads on the custom CI infrastructure.
+
+## Project Layout and Architecture
+
+### Core Directories
+- **`src/`**: Main llama library implementation (`llama.cpp`, `llama-*.cpp`)
+- **`include/`**: Public API headers, primarily `include/llama.h`
+- **`ggml/`**: Core tensor library (submodule with custom GGML framework)
+- **`examples/`**: 30+ example applications and tools
+- **`tools/`**: Additional development and utility tools (server benchmarks, tests)
+- **`tests/`**: Comprehensive test suite with CTest integration
+- **`docs/`**: Detailed documentation (build guides, API docs, etc.)
+- **`scripts/`**: Utility scripts for CI, data processing, and automation
+- **`common/`**: Shared utility code used across examples
+
+### Key Files
+- **`CMakeLists.txt`**: Primary build configuration
+- **`include/llama.h`**: Main C API header (~2000 lines)
+- **`src/llama.cpp`**: Core library implementation (~8000 lines)
+- **`CONTRIBUTING.md`**: Coding guidelines and PR requirements
+- **`.clang-format`**: C++ formatting rules
+- **`.pre-commit-config.yaml`**: Git hook configuration
+
+### Built Executables (in `build/bin/`)
+Primary tools:
+- **`llama-cli`**: Main inference tool
+- **`llama-server`**: OpenAI-compatible HTTP server
+- **`llama-quantize`**: Model quantization utility
+- **`llama-perplexity`**: Model evaluation tool
+- **`llama-bench`**: Performance benchmarking
+- **`llama-convert-llama2c-to-ggml`**: Model conversion utilities
+
+### Configuration Files
+- **CMake**: `CMakeLists.txt`, `cmake/` directory
+- **Linting**: `.clang-format`, `.clang-tidy`, `.flake8`
+- **CI**: `.github/workflows/`, `ci/run.sh`
+- **Git**: `.gitignore` (includes build artifacts, models, cache)
+
+### Dependencies
+- **System**: OpenMP, libcurl (for model downloading)
+- **Optional**: CUDA SDK, Metal framework, Vulkan SDK, Intel oneAPI
+- **Bundled**: httplib, json (header-only libraries in vendored form)
+
+## Common Validation Steps
+
+### After Making Changes
+1. **Format code**: `git clang-format`
+2. **Build**: `cmake --build build --config Release`
+3. **Test**: `ctest --test-dir build --output-on-failure`
+4. **Server tests** (if modifying server): `cd tools/server/tests && source ../../../.venv/bin/activate && ./tests.sh`
+5. **Manual validation**: Test relevant tools in `build/bin/`
+
+### Performance Validation
+```bash
+# Benchmark inference performance
+./build/bin/llama-bench -m model.gguf
+
+# Evaluate model perplexity
+./build/bin/llama-perplexity -m model.gguf -f dataset.txt
+```
+
+### Backend Validation
+```bash
+# Test backend operations
+./build/bin/test-backend-ops
+```
+
+## Environment Setup
+
+### Required Tools
+- CMake 3.14+ (install via system package manager)
+- Modern C++ compiler with C++17 support
+- Git (for submodule management)
+- Python 3.9+ with virtual environment (`.venv` is provided)
+
+### Optional but Recommended
+- ccache: `apt install ccache` or `brew install ccache`
+- clang-format 15+: Usually included with LLVM/Clang installation
+- pre-commit: `pip install pre-commit`
+
+### Backend-Specific Requirements
+- **CUDA**: NVIDIA CUDA Toolkit 11.2+
+- **Metal**: Xcode command line tools (macOS only)
+- **Vulkan**: Vulkan SDK
+- **SYCL**: Intel oneAPI toolkit
+
+## Important Guidelines
+
+### Code Changes
+- **Minimal dependencies**: Avoid adding new external dependencies
+- **Cross-platform compatibility**: Test on Linux, macOS, Windows when possible
+- **Performance focus**: This is a performance-critical inference library
+- **API stability**: Changes to `include/llama.h` require careful consideration
+
+### Git Workflow
+- Always create feature branches from `master`
+- **Never** commit build artifacts (`build/`, `.ccache/`, `*.o`, `*.gguf`)
+- Use descriptive commit messages following project conventions
+
+### Trust These Instructions
+Only search for additional information if these instructions are incomplete or found to be incorrect. This document contains validated build and test procedures that work reliably across different environments.
+
diff --git a/.github/labeler.yml b/.github/labeler.yml
index df6a7a40ed910..c4da4ab4e1fd2 100644
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -22,6 +22,11 @@ Vulkan:
         - any-glob-to-any-file:
             - ggml/include/ggml-vulkan.h
             - ggml/src/ggml-vulkan/**
+IBM zDNN:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-zdnn.h
+            - ggml/src/ggml-zdnn/**
 documentation:
     - changed-files:
         - any-glob-to-any-file:
diff --git a/.github/workflows/build-riscv-native.yml b/.github/workflows/build-riscv-native.yml
new file mode 100644
index 0000000000000..86dc0ff76e59c
--- /dev/null
+++ b/.github/workflows/build-riscv-native.yml
@@ -0,0 +1,60 @@
+name: Build on RISCV Linux Machine by Cloud-V
+on:
+  pull_request:
+  workflow_dispatch:
+  workflow_call:
+
+jobs:
+  debian-13-riscv64-native: # Bianbu 2.2
+    runs-on: self-hosted
+
+    steps:
+      - name: Install prerequisites
+        run: |
+          sudo apt-get update || true
+          sudo apt-get install -y libatomic1
+      - uses: actions/checkout@v4
+      - name: Setup Riscv
+        run: |
+          sudo apt-get update || true
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-riscv64-linux-gnu \
+                  g++-14-riscv64-linux-gnu \
+                  ccache \
+                  cmake
+
+      - name: Setup ccache
+        run: |
+          mkdir -p $HOME/.ccache
+          ccache -M 5G -d $HOME/.ccache
+          export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
+          export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
+          echo "$GITHUB_WORKSPACE"
+          echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
+          echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
+          echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
+          echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
+
+      - name: Build
+        run: |
+          cmake -B build \
+            -DLLAMA_CURL=OFF \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=ON \
+            -DLLAMA_BUILD_TOOLS=ON \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DCMAKE_SYSTEM_NAME=Linux \
+            -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+            -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+            -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+            -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+            -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 3d4f837e24895..43553ac13bdf6 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -64,7 +64,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
@@ -104,7 +104,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
@@ -144,7 +144,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-arm64-webgpu
           evict-old-files: 1d
@@ -199,7 +199,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-cpu-cmake
           evict-old-files: 1d
@@ -251,7 +251,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-latest-cmake-sanitizer-${{ matrix.sanitizer }}
           evict-old-files: 1d
@@ -330,7 +330,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-latest-cmake-rpc
           evict-old-files: 1d
@@ -363,7 +363,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-vulkan
           evict-old-files: 1d
@@ -400,7 +400,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-webgpu
           evict-old-files: 1d
@@ -443,7 +443,7 @@ jobs:
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
-    container: rocm/dev-ubuntu-22.04:6.0.2
+    container: rocm/dev-ubuntu-22.04:6.1.2
 
     steps:
       - name: Clone
@@ -457,7 +457,7 @@ jobs:
           sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-hip
           evict-old-files: 1d
@@ -471,16 +471,6 @@ jobs:
             -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
-      - name: Build with legacy HIP support
-        id: cmake_build_legacy_hip
-        run: |
-          cmake -B build2 -S . \
-            -DCMAKE_C_COMPILER=hipcc \
-            -DCMAKE_CXX_COMPILER=hipcc \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
-            -DGGML_HIP=ON
-          cmake --build build2 --config Release -j $(nproc)
-
   ubuntu-22-cmake-musa:
     runs-on: ubuntu-22.04
     container: mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
@@ -497,7 +487,7 @@ jobs:
           apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-musa
           evict-old-files: 1d
@@ -542,7 +532,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-sycl
           evict-old-files: 1d
@@ -590,7 +580,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-sycl-fp16
           evict-old-files: 1d
@@ -621,7 +611,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-ios
           evict-old-files: 1d
@@ -658,7 +648,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-tvos
           evict-old-files: 1d
@@ -730,7 +720,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-swift
           evict-old-files: 1d
@@ -776,7 +766,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-msys2
           variant: ccache
@@ -844,7 +834,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-${{ matrix.build }}
           variant: ccache
@@ -958,7 +948,7 @@ jobs:
               apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev
 
         - name: ccache
-          uses: hendrikmuhs/ccache-action@v1.2.16
+          uses: ggml-org/ccache-action@v1.2.16
           with:
             key: ubuntu-latest-cmake-cuda
             evict-old-files: 1d
@@ -987,7 +977,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-cuda-${{ matrix.cuda }}
           variant: ccache
@@ -1043,7 +1033,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-sycl
           variant: ccache
@@ -1080,7 +1070,8 @@ jobs:
           write-host "Downloading AMD HIP SDK Installer"
           Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
-          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
+          $proc.WaitForExit(600000)
           write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
@@ -1089,7 +1080,7 @@ jobs:
           & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
 
       - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ${{ github.job }}
           evict-old-files: 1d
@@ -1123,6 +1114,11 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
+      - name: Setup Xcode
+        uses: maxim-lobanov/setup-xcode@v1
+        with:
+          xcode-version: latest-stable
+
       - name: Build
         id: cmake_build
         run: |
@@ -1156,7 +1152,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: android-build
           evict-old-files: 1d
diff --git a/.github/workflows/copilot-setup-steps.yml b/.github/workflows/copilot-setup-steps.yml
new file mode 100644
index 0000000000000..3645e30378b95
--- /dev/null
+++ b/.github/workflows/copilot-setup-steps.yml
@@ -0,0 +1,57 @@
+name: "Copilot Setup Steps"
+
+# Automatically run the setup steps when they are changed to allow for easy validation, and
+# allow manual testing through the repository's "Actions" tab
+on:
+  workflow_dispatch:
+  push:
+    paths:
+      - .github/workflows/copilot-setup-steps.yml
+  pull_request:
+    paths:
+      - .github/workflows/copilot-setup-steps.yml
+
+jobs:
+  # The job MUST be called `copilot-setup-steps` or it will not be picked up by Copilot.
+  copilot-setup-steps:
+    runs-on: ubuntu-latest
+
+    # Set the permissions to the lowest permissions possible needed for your steps.
+    # Copilot will be given its own token for its operations.
+    permissions:
+      # If you want to clone the repository as part of your setup steps, for example to install dependencies, you'll need the `contents: read` permission. If you don't clone the repository in your setup steps, Copilot will do this for you automatically after the steps complete.
+      contents: read
+
+    # You can define any steps you want, and they will run before the agent starts.
+    # If you do not check out your code, Copilot will do this for you.
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: copilot-setup-steps
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+          # Install git-clang-format script for formatting only changed code
+          wget -O /tmp/git-clang-format https://raw.githubusercontent.com/llvm/llvm-project/release/18.x/clang/tools/clang-format/git-clang-format
+          sudo cp /tmp/git-clang-format /usr/local/bin/git-clang-format
+          sudo chmod +x /usr/local/bin/git-clang-format
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Install Python dependencies
+        run: |
+          python3 -m venv .venv
+          .venv/bin/activate
+          pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
+          pip install flake8 pyright pre-commit
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 4ed6126f487c0..5367637e42843 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -32,7 +32,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
@@ -85,7 +85,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
@@ -147,7 +147,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-cpu-cmake
           evict-old-files: 1d
@@ -198,7 +198,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-vulkan
           evict-old-files: 1d
@@ -256,7 +256,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-cpu-${{ matrix.arch }}
           variant: ccache
@@ -328,7 +328,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
           variant: ccache
@@ -398,7 +398,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-cuda-${{ matrix.cuda }}
           variant: ccache
@@ -471,7 +471,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-sycl
           variant: ccache
@@ -545,7 +545,7 @@ jobs:
           git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-hip-${{ matrix.name }}-x64
           evict-old-files: 1d
@@ -557,7 +557,8 @@ jobs:
           write-host "Downloading AMD HIP SDK Installer"
           Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
-          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
+          $proc.WaitForExit(600000)
           write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
@@ -600,7 +601,7 @@ jobs:
           name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
 
   ios-xcode-build:
-    runs-on: macos-latest
+    runs-on: macos-15
 
     steps:
       - name: Checkout code
@@ -608,6 +609,10 @@ jobs:
         with:
           fetch-depth: 0
 
+      - name: Setup Xcode
+        run: |
+          sudo xcode-select -s /Applications/Xcode_16.4.app
+
       - name: Build
         id: cmake_build
         run: |
diff --git a/.gitignore b/.gitignore
index f48ce4cacd144..595831accb05d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -147,3 +147,4 @@ poetry.toml
 # Local scripts
 /run-vim.sh
 /run-chat.sh
+.ccache/
diff --git a/CMakeLists.txt b/CMakeLists.txt
index c79ccd09e097c..36a2078e4c9fa 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -12,6 +12,8 @@ if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
     set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
 endif()
 
+message("CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}")
+
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
 
diff --git a/CODEOWNERS b/CODEOWNERS
index 4c0dd4b725dd1..18564a08b1d6c 100644
--- a/CODEOWNERS
+++ b/CODEOWNERS
@@ -5,8 +5,8 @@
 /tools/server/ @ngxson
 /ggml/src/ggml-cuda/fattn* @JohannesGaessler
 /ggml/src/ggml-cuda/mmq.* @JohannesGaessler
-/ggml/src/ggml-cuda/mmv.* @JohannesGaessler
 /ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
 /ggml/src/ggml-opt.cpp @JohannesGaessler
 /ggml/src/gguf.cpp @JohannesGaessler
 /ggml/src/ggml-vulkan/ @0cc4m
+/ggml/src/ggml-zdnn/ @taronaeo
diff --git a/Makefile b/Makefile
index ac442aec095d6..bcbc770205956 100644
--- a/Makefile
+++ b/Makefile
@@ -1,1608 +1,9 @@
-ifndef LLAMA_MAKEFILE
-$(error The Makefile build is deprecated. Use the CMake build instead. For more details, see https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
-endif
+define newline
 
-# Define the default target now so that it is always the first target
-BUILD_TARGETS = \
-	libllava.a \
-	llama-batched \
-	llama-batched-bench \
-	llama-bench \
-	llama-cli \
-	llama-convert-llama2c-to-ggml \
-	llama-embedding \
-	llama-eval-callback \
-	llama-export-lora \
-	llama-gbnf-validator \
-	llama-gguf \
-	llama-gguf-hash \
-	llama-gguf-split \
-	llama-gritlm \
-	llama-imatrix \
-	llama-infill \
-	llama-llava-cli \
-	llama-minicpmv-cli\
-	llama-qwen2vl-cli\
-	llama-lookahead \
-	llama-lookup \
-	llama-lookup-create \
-	llama-lookup-merge \
-	llama-lookup-stats \
-	llama-parallel \
-	llama-passkey \
-	llama-perplexity \
-	llama-q8dot \
-	llama-quantize \
-	llama-quantize-stats \
-	llama-retrieval \
-	llama-save-load-state \
-	llama-server \
-	llama-simple \
-	llama-simple-chat \
-	llama-run \
-	llama-speculative \
-	llama-tokenize \
-	llama-vdot \
-	llama-cvector-generator \
-	llama-gen-docs \
-	tests/test-c.o
 
-# Binaries only useful for tests
-TEST_TARGETS = \
-	tests/test-arg-parser \
-	tests/test-autorelease \
-	tests/test-backend-ops \
-	tests/test-chat \
-	tests/test-chat-template \
-	tests/test-double-float \
-	tests/test-grammar-integration \
-	tests/test-grammar-parser \
-	tests/test-json-schema-to-grammar \
-	tests/test-llama-grammar \
-	tests/test-log \
-	tests/test-model-load-cancel \
-	tests/test-quantize-fns \
-	tests/test-quantize-perf \
-	tests/test-rope \
-	tests/test-sampling \
-	tests/test-tokenizer-0 \
-	tests/test-tokenizer-1-bpe \
-	tests/test-tokenizer-1-spm
-#	tests/test-opt \
+endef
 
-# Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
-LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \
-	retrieval speculative infill tokenize parallel export-lora lookahead lookup passkey gritlm
-
-# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them.
-#  We don't want to clutter things too much, so we only build replacements for the most commonly used binaries.
-LEGACY_TARGETS_BUILD = main quantize perplexity embedding server
-
-# Deprecation aliases
-ifdef LLAMA_CUBLAS
-$(error LLAMA_CUBLAS is removed. Use GGML_CUDA instead.)
-endif
-
-ifdef LLAMA_CUDA
-GGML_CUDA := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_KOMPUTE
-GGML_KOMPUTE := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_METAL
-GGML_METAL := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_RPC
-GGML_RPC := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_SYCL
-GGML_SYCL := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_SYCL_F16
-GGML_SYCL_F16 := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_OPENBLAS
-GGML_OPENBLAS := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_OPENBLAS64
-GGML_OPENBLAS64 := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_BLIS
-GGML_BLIS := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_NO_LLAMAFILE
-GGML_NO_LLAMAFILE := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_NO_ACCELERATE
-GGML_NO_ACCELERATE := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_NO_OPENMP
-GGML_NO_OPENMP := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_NO_METAL
-GGML_NO_METAL := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef LLAMA_DISABLE_LOGS
-REMOVE_WARNING := 1
-endif
-
-ifdef LLAMA_SERVER_VERBOSE
-REMOVE_WARNING := 1
-endif
-
-ifndef UNAME_S
-UNAME_S := $(shell uname -s)
-endif
-
-ifndef UNAME_P
-UNAME_P := $(shell uname -p)
-endif
-
-ifndef UNAME_M
-UNAME_M := $(shell uname -m)
-endif
-
-# In GNU make default CXX is g++ instead of c++.  Let's fix that so that users
-# of non-gcc compilers don't have to provide g++ alias or wrapper.
-DEFCC  := cc
-DEFCXX := c++
-ifeq ($(origin CC),default)
-CC  := $(DEFCC)
-endif
-ifeq ($(origin CXX),default)
-CXX := $(DEFCXX)
-endif
-
-# Mac OS + Arm can report x86_64
-# ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
-ifeq ($(UNAME_S),Darwin)
-	ifndef GGML_NO_METAL
-		GGML_METAL := 1
-	endif
-
-	GGML_NO_OPENMP := 1
-
-	ifneq ($(UNAME_P),arm)
-		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
-		ifeq ($(SYSCTL_M),1)
-			# UNAME_P := arm
-			# UNAME_M := arm64
-			warn := $(warning Your arch is announced as x86_64, but it seems to actually be ARM64. Not fixing that can lead to bad performance. For more info see: https://github.com/ggerganov/whisper.cpp/issues/66\#issuecomment-1282546789)
-		endif
-	endif
-endif
-
-ifdef GGML_METAL
-	GGML_METAL_EMBED_LIBRARY := 1
-endif
-
-ifdef GGML_RPC
-	BUILD_TARGETS += rpc-server
-endif
-
-ifdef GGML_VULKAN
-	BUILD_TARGETS += vulkan-shaders-gen
-endif
-
-default: $(BUILD_TARGETS) $(LEGACY_TARGETS_BUILD)
-
-test: $(TEST_TARGETS)
-	@failures=0; \
-	for test_target in $(TEST_TARGETS); do \
-		if [ "$$test_target" = "tests/test-tokenizer-0" ]; then \
-			./$$test_target $(CURDIR)/models/ggml-vocab-llama-spm.gguf; \
-			./$$test_target $(CURDIR)/models/ggml-vocab-llama-bpe.gguf; \
-			./$$test_target $(CURDIR)/models/ggml-vocab-phi-3.gguf; \
-			./$$test_target $(CURDIR)/models/ggml-vocab-falcon.gguf; \
-			./$$test_target $(CURDIR)/models/ggml-vocab-bert-bge.gguf; \
-			./$$test_target $(CURDIR)/models/ggml-vocab-starcoder.gguf; \
-			./$$test_target $(CURDIR)/models/ggml-vocab-gpt-2.gguf; \
-			./$$test_target $(CURDIR)/models/ggml-vocab-refact.gguf; \
-		elif [ "$$test_target" = "tests/test-tokenizer-1-spm" ]; then \
-			continue; \
-		elif [ "$$test_target" = "tests/test-tokenizer-1-bpe" ]; then \
-			continue; \
-		else \
-			echo "Running test $$test_target..."; \
-			./$$test_target; \
-		fi; \
-		if [ $$? -ne 0 ]; then \
-			printf 'Test %s FAILED!\n\n' $$test_target; \
-			failures=$$(( failures + 1 )); \
-		else \
-			printf 'Test %s passed.\n\n' $$test_target; \
-		fi; \
-	done; \
-	if [ $$failures -gt 0 ]; then \
-		printf '\n%s tests failed.\n' $$failures; \
-		exit 1; \
-	fi
-	@echo 'All tests passed.'
-
-all: $(BUILD_TARGETS) $(TEST_TARGETS) $(LEGACY_TARGETS_BUILD)
-
-ifdef RISCV_CROSS_COMPILE
-CC	:= riscv64-unknown-linux-gnu-gcc
-CXX	:= riscv64-unknown-linux-gnu-g++
-endif
-
-#
-# Compile flags
-#
-
-# keep standard at C11 and C++17
-MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon -DGGML_USE_CPU
-MK_CFLAGS    = -std=c11   -fPIC
-MK_CXXFLAGS  = -std=c++17 -fPIC
-MK_NVCCFLAGS = -std=c++17
-
-ifdef LLAMA_NO_CCACHE
-GGML_NO_CCACHE := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifndef GGML_NO_CCACHE
-CCACHE := $(shell which ccache)
-ifdef CCACHE
-export CCACHE_SLOPPINESS = time_macros
-$(info I ccache found, compilation results will be cached. Disable with GGML_NO_CCACHE.)
-CC    := $(CCACHE) $(CC)
-CXX   := $(CCACHE) $(CXX)
-else
-$(info I ccache not found. Consider installing it for faster compilation.)
-endif # CCACHE
-endif # GGML_NO_CCACHE
-
-# clock_gettime came in POSIX.1b (1993)
-# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
-# posix_memalign came in POSIX.1-2001 / SUSv3
-# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-MK_CPPFLAGS += -D_XOPEN_SOURCE=600
-
-# Somehow in OpenBSD whenever POSIX conformance is specified
-# some string functions rely on locale_t availability,
-# which was introduced in POSIX.1-2008, forcing us to go higher
-ifeq ($(UNAME_S),OpenBSD)
-	MK_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
-endif
-
-# Data types, macros and functions related to controlling CPU affinity and
-# some memory allocation are available on Linux through GNU extensions in libc
-ifeq ($(UNAME_S),Linux)
-	MK_CPPFLAGS += -D_GNU_SOURCE
-	MK_LDFLAGS  += -ldl
-endif
-
-# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
-# and on macOS its availability depends on enabling Darwin extensions
-# similarly on DragonFly, enabling BSD extensions is necessary
-ifeq ($(UNAME_S),Darwin)
-	MK_CPPFLAGS += -D_DARWIN_C_SOURCE
-endif
-ifeq ($(UNAME_S),DragonFly)
-	MK_CPPFLAGS += -D__BSD_VISIBLE
-endif
-
-# alloca is a non-standard interface that is not visible on BSDs when
-# POSIX conformance is specified, but not all of them provide a clean way
-# to enable it in such cases
-ifeq ($(UNAME_S),FreeBSD)
-	MK_CPPFLAGS += -D__BSD_VISIBLE
-endif
-ifeq ($(UNAME_S),NetBSD)
-	MK_CPPFLAGS += -D_NETBSD_SOURCE
-endif
-ifeq ($(UNAME_S),OpenBSD)
-	MK_CPPFLAGS += -D_BSD_SOURCE
-endif
-
-ifdef GGML_SCHED_MAX_COPIES
-	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(GGML_SCHED_MAX_COPIES)
-endif
-
-ifdef LLAMA_DEBUG
-	MK_CFLAGS    += -O0 -g
-	MK_CXXFLAGS  += -O0 -g
-	MK_LDFLAGS   += -g
-	MK_NVCCFLAGS += -O0 -g
-
-	ifeq ($(UNAME_S),Linux)
-		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
-	endif
-else
-	MK_CPPFLAGS   += -DNDEBUG
-	MK_CFLAGS     += -O3 -g
-	MK_CXXFLAGS   += -O3 -g
-	MK_NVCCFLAGS  += -O3 -g
-endif
-
-ifdef LLAMA_SANITIZE_THREAD
-	MK_CFLAGS   += -fsanitize=thread -g
-	MK_CXXFLAGS += -fsanitize=thread -g
-	MK_LDFLAGS  += -fsanitize=thread -g
-endif
-
-ifdef LLAMA_SANITIZE_ADDRESS
-	MK_CFLAGS   += -fsanitize=address -fno-omit-frame-pointer -g
-	MK_CXXFLAGS += -fsanitize=address -fno-omit-frame-pointer -g
-	MK_LDFLAGS  += -fsanitize=address -fno-omit-frame-pointer -g
-endif
-
-ifdef LLAMA_SANITIZE_UNDEFINED
-	MK_CFLAGS   += -fsanitize=undefined -g
-	MK_CXXFLAGS += -fsanitize=undefined -g
-	MK_LDFLAGS  += -fsanitize=undefined -g
-endif
-
-ifdef LLAMA_SERVER_SSL
-	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
-	MK_LDFLAGS += -lssl -lcrypto
-endif
-
-ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_REPACK
-endif
-
-# warnings
-WARN_FLAGS = \
-	-Wall \
-	-Wextra \
-	-Wpedantic \
-	-Wcast-qual \
-	-Wno-unused-function
-
-MK_CFLAGS += \
-	$(WARN_FLAGS) \
-	-Wshadow \
-	-Wstrict-prototypes \
-	-Wpointer-arith \
-	-Wmissing-prototypes \
-	-Werror=implicit-int \
-	-Werror=implicit-function-declaration
-
-MK_CXXFLAGS += \
-	$(WARN_FLAGS) \
-	-Wmissing-declarations \
-	-Wmissing-noreturn
-
-ifeq ($(LLAMA_FATAL_WARNINGS),1)
-	MK_CFLAGS   += -Werror
-	MK_CXXFLAGS += -Werror
-endif
-
-# this version of Apple ld64 is buggy
-ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
-	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
-endif
-
-# OS specific
-# TODO: support Windows
-ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)'
-	MK_CFLAGS   += -pthread
-	MK_CXXFLAGS += -pthread
-endif
-
-# detect Windows
-ifneq ($(findstring _NT,$(UNAME_S)),)
-	_WIN32 := 1
-endif
-
-# library name prefix
-ifneq ($(_WIN32),1)
-	LIB_PRE := lib
-endif
-
-# Dynamic Shared Object extension
-ifneq ($(_WIN32),1)
-	DSO_EXT := .so
-else
-	DSO_EXT := .dll
-endif
-
-# Windows Sockets 2 (Winsock) for network-capable apps
-ifeq ($(_WIN32),1)
-	LWINSOCK2 := -lws2_32
-endif
-
-ifdef LLAMA_GPROF
-	MK_CFLAGS   += -pg
-	MK_CXXFLAGS += -pg
-endif
-
-# Architecture specific
-# TODO: probably these flags need to be tweaked on some architectures
-#       feel free to update the Makefile for your architecture and send a pull request or issue
-
-ifndef RISCV_CROSS_COMPILE
-
-ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
-	# Use all CPU extensions that are available:
-	MK_CFLAGS     += -march=native -mtune=native
-	HOST_CXXFLAGS += -march=native -mtune=native
-
-	# Usage AMX build test
-	#MK_CFLAGS     += -march=graniterapids -mtune=graniterapids
-	#HOST_CXXFLAGS += -march=graniterapids -mtune=graniterapids
-
-	# Usage AVX-only
-	#MK_CFLAGS   += -mfma -mf16c -mavx
-	#MK_CXXFLAGS += -mfma -mf16c -mavx
-
-	# Usage SSSE3-only (Not is SSE3!)
-	#MK_CFLAGS   += -mssse3
-	#MK_CXXFLAGS += -mssse3
-endif
-
-ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
-	# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
-	# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
-	# https://github.com/ggml-org/llama.cpp/issues/2922
-	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
-	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
-
-	# Target Windows 8 for PrefetchVirtualMemory
-	MK_CPPFLAGS += -D_WIN32_WINNT=0x602
-endif
-
-ifneq ($(filter aarch64%,$(UNAME_M)),)
-	# Apple M1, M2, etc.
-	# Raspberry Pi 3, 4, Zero 2 (64-bit)
-	# Nvidia Jetson
-	MK_CFLAGS   += -mcpu=native
-	MK_CXXFLAGS += -mcpu=native
-	JETSON_RELEASE_INFO = $(shell jetson_release)
-	ifdef JETSON_RELEASE_INFO
-		ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),)
-			JETSON_EOL_MODULE_DETECT = 1
-			CC = aarch64-unknown-linux-gnu-gcc
-			cxx = aarch64-unknown-linux-gnu-g++
-		endif
-	endif
-endif
-
-ifneq ($(filter armv6%,$(UNAME_M)),)
-	# Raspberry Pi 1, Zero
-	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
-	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
-endif
-
-ifneq ($(filter armv7%,$(UNAME_M)),)
-	# Raspberry Pi 2
-	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
-	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
-endif
-
-ifneq ($(filter armv8%,$(UNAME_M)),)
-	# Raspberry Pi 3, 4, Zero 2 (32-bit)
-	MK_CFLAGS   += -mfp16-format=ieee -mno-unaligned-access
-	MK_CXXFLAGS += -mfp16-format=ieee -mno-unaligned-access
-endif
-
-ifneq ($(filter ppc64%,$(UNAME_M)),)
-	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
-	ifneq (,$(findstring POWER9,$(POWER9_M)))
-		MK_CFLAGS   += -mcpu=power9
-		MK_CXXFLAGS += -mcpu=power9
-	endif
-endif
-
-ifneq ($(filter ppc64le%,$(UNAME_M)),)
-	MK_CFLAGS   += -mcpu=powerpc64le
-	MK_CXXFLAGS += -mcpu=powerpc64le
-	CUDA_POWER_ARCH = 1
-endif
-
-ifneq ($(filter loongarch64%,$(UNAME_M)),)
-	MK_CFLAGS   += -mlasx
-	MK_CXXFLAGS += -mlasx
-endif
-
-ifneq ($(filter riscv64%,$(UNAME_M)),)
-	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
-	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
-endif
-
-else # RISC-V CROSS COMPILATION
-	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
-	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
-endif
-
-ifndef GGML_NO_ACCELERATE
-	# Mac OS - include Accelerate framework.
-	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
-	ifeq ($(UNAME_S),Darwin)
-		MK_CPPFLAGS  += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE
-		MK_CPPFLAGS  += -DACCELERATE_NEW_LAPACK
-		MK_CPPFLAGS  += -DACCELERATE_LAPACK_ILP64
-		MK_LDFLAGS   += -framework Accelerate
-		OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
-	endif
-endif # GGML_NO_ACCELERATE
-
-ifndef GGML_NO_OPENMP
-	MK_CPPFLAGS += -DGGML_USE_OPENMP
-	MK_CFLAGS   += -fopenmp
-	MK_CXXFLAGS += -fopenmp
-endif # GGML_NO_OPENMP
-
-ifdef GGML_OPENBLAS
-	MK_CPPFLAGS  += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
-	MK_CFLAGS    += $(shell pkg-config --cflags-only-other openblas)
-	MK_LDFLAGS   += $(shell pkg-config --libs openblas)
-	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
-endif # GGML_OPENBLAS
-
-ifdef GGML_OPENBLAS64
-	MK_CPPFLAGS  += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
-	MK_CFLAGS    += $(shell pkg-config --cflags-only-other openblas64)
-	MK_LDFLAGS   += $(shell pkg-config --libs openblas64)
-	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
-endif # GGML_OPENBLAS64
-
-ifdef GGML_BLIS
-	MK_CPPFLAGS  += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
-	MK_LDFLAGS   += -lblis -L/usr/local/lib
-	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
-endif # GGML_BLIS
-
-ifdef GGML_NVPL
-	MK_CPPFLAGS  += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
-	MK_LDFLAGS   += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
-	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
-endif # GGML_NVPL
-
-ifndef GGML_NO_LLAMAFILE
-	MK_CPPFLAGS  += -DGGML_USE_LLAMAFILE
-	OBJ_GGML_EXT += ggml/src/ggml-cpu/llamafile/sgemm.o
-endif
-
-ifndef GGML_NO_AMX
-	MK_CPPFLAGS += -DGGML_USE_AMX
-	OBJ_GGML_EXT += ggml/src/ggml-cpu/amx/amx.o ggml/src/ggml-cpu/amx/mmq.o
-endif
-
-# only necessary for the CPU backend files
-MK_CPPFLAGS += -Iggml/src/ggml-cpu
-
-ifdef GGML_RPC
-	MK_CPPFLAGS  += -DGGML_USE_RPC
-	OBJ_GGML_EXT += ggml/src/ggml-rpc.o
-endif # GGML_RPC
-
-OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-mma*.cu))
-OBJ_CUDA_TMPL     += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/mmq*.cu))
-
-ifdef GGML_CUDA_FA_ALL_QUANTS
-	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*.cu))
-else
-	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu))
-	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu))
-	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*f16-f16.cu))
-endif # GGML_CUDA_FA_ALL_QUANTS
-
-ifdef GGML_CUDA
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
-	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
-
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-	MK_NVCCFLAGS += -use_fast_math
-
-	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
-
-ifdef LLAMA_FATAL_WARNINGS
-	MK_NVCCFLAGS += -Werror all-warnings
-endif # LLAMA_FATAL_WARNINGS
-
-ifndef JETSON_EOL_MODULE_DETECT
-	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
-endif # JETSON_EOL_MODULE_DETECT
-
-ifdef LLAMA_DEBUG
-	MK_NVCCFLAGS += -lineinfo
-endif # LLAMA_DEBUG
-
-ifdef GGML_CUDA_DEBUG
-	MK_NVCCFLAGS += --device-debug
-endif # GGML_CUDA_DEBUG
-
-ifdef GGML_CUDA_NVCC
-	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
-else
-	NVCC = $(CCACHE) nvcc
-endif # GGML_CUDA_NVCC
-
-ifdef CUDA_DOCKER_ARCH
-	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
-else ifndef CUDA_POWER_ARCH
-	MK_NVCCFLAGS += -arch=native
-endif # CUDA_DOCKER_ARCH
-
-ifdef GGML_CUDA_FORCE_MMQ
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
-endif # GGML_CUDA_FORCE_MMQ
-
-ifdef GGML_CUDA_FORCE_CUBLAS
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_CUBLAS
-endif # GGML_CUDA_FORCE_CUBLAS
-
-ifdef GGML_CUDA_F16
-	MK_NVCCFLAGS += -DGGML_CUDA_F16
-endif # GGML_CUDA_F16
-
-ifdef GGML_CUDA_DMMV_F16
-	MK_NVCCFLAGS += -DGGML_CUDA_F16
-endif # GGML_CUDA_DMMV_F16
-
-ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
-else
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
-endif # GGML_CUDA_PEER_MAX_BATCH_SIZE
-
-ifdef GGML_CUDA_NO_PEER_COPY
-	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
-endif # GGML_CUDA_NO_PEER_COPY
-
-ifdef GGML_CUDA_CCBIN
-	MK_NVCCFLAGS += -ccbin $(GGML_CUDA_CCBIN)
-endif # GGML_CUDA_CCBIN
-
-ifdef GGML_CUDA_NO_FA
-	MK_NVCCFLAGS += -DGGML_CUDA_NO_FA
-endif # GGML_CUDA_NO_FA
-
-ifdef GGML_CUDA_FA_ALL_QUANTS
-	MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
-endif # GGML_CUDA_FA_ALL_QUANTS
-
-ifdef JETSON_EOL_MODULE_DETECT
-define NVCC_COMPILE
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
-endef # NVCC_COMPILE
-else
-define NVCC_COMPILE
-	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
-endef # NVCC_COMPILE
-endif # JETSON_EOL_MODULE_DETECT
-
-ggml/src/ggml-cuda/%.o: \
-	ggml/src/ggml-cuda/%.cu \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h \
-	ggml/src/ggml-cuda/common.cuh
-	$(NVCC_COMPILE)
-
-ggml/src/ggml-cuda/ggml-cuda.o: \
-	ggml/src/ggml-cuda/ggml-cuda.cu \
-	ggml/include/ggml-cuda.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h \
-	ggml/src/ggml-backend-impl.h \
-	ggml/src/ggml-common.h \
-	$(wildcard ggml/src/ggml-cuda/*.cuh)
-	$(NVCC_COMPILE)
-endif # GGML_CUDA
-
-ifdef GGML_VULKAN
-	MK_CPPFLAGS  += -DGGML_USE_VULKAN
-	MK_LDFLAGS   += $(shell pkg-config --libs vulkan)
-	OBJ_GGML_EXT += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
-
-ifdef GGML_VULKAN_CHECK_RESULTS
-	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
-endif
-
-ifdef GGML_VULKAN_DEBUG
-	MK_CPPFLAGS  += -DGGML_VULKAN_DEBUG
-endif
-
-ifdef GGML_VULKAN_MEMORY_DEBUG
-	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
-endif
-
-ifdef GGML_VULKAN_PERF
-	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
-endif
-
-ifdef GGML_VULKAN_VALIDATE
-	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
-endif
-
-ifdef GGML_VULKAN_RUN_TESTS
-	MK_CPPFLAGS  += -DGGML_VULKAN_RUN_TESTS
-endif
-
-GLSLC_CMD  = glslc
-_ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen
-_ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp
-_ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp
-_ggml_vk_input_dir = ggml/src/ggml-vulkan/vulkan-shaders
-_ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp)
-
-ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
-	$(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@
-
-$(_ggml_vk_header): $(_ggml_vk_source)
-
-$(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen
-	$(_ggml_vk_genshaders_cmd) \
-		--glslc      $(GLSLC_CMD) \
-		--input-dir  $(_ggml_vk_input_dir) \
-		--target-hpp $(_ggml_vk_header) \
-		--target-cpp $(_ggml_vk_source)
-
-vulkan-shaders-gen: ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
-
-endif # GGML_VULKAN
-
-ifdef GGML_HIP
-	ifeq ($(wildcard /opt/rocm),)
-		ROCM_PATH      ?= /usr
-		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
-	else
-		ROCM_PATH	?= /opt/rocm
-		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
-	endif
-
-	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
-
-	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
-	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
-	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas
-
-	HIPCC ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
-
-	HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
-
-ifdef GGML_CUDA_FORCE_MMQ
-	HIPFLAGS += -DGGML_CUDA_FORCE_MMQ
-endif # GGML_CUDA_FORCE_MMQ
-
-ifdef GGML_CUDA_FORCE_CUBLAS
-	HIPFLAGS += -DGGML_CUDA_FORCE_CUBLAS
-endif # GGML_CUDA_FORCE_CUBLAS
-
-ifdef GGML_CUDA_NO_PEER_COPY
-	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
-endif # GGML_CUDA_NO_PEER_COPY
-
-ifdef GGML_CUDA_NO_FA
-	HIPFLAGS += -DGGML_CUDA_NO_FA
-endif # GGML_CUDA_NO_FA
-
-	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
-
-ggml/src/ggml-cuda/ggml-cuda.o: \
-	ggml/src/ggml-cuda/ggml-cuda.cu \
-	ggml/include/ggml-cuda.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h \
-	ggml/src/ggml-backend-impl.h \
-	ggml/src/ggml-common.h \
-	$(wildcard ggml/src/ggml-cuda/*.cuh)
-	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
-
-ggml/src/ggml-cuda/%.o: \
-	ggml/src/ggml-cuda/%.cu \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h \
-	ggml/src/ggml-cuda/common.cuh
-	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
-endif # GGML_HIP
-
-ifdef GGML_MUSA
-	ifeq ($(wildcard /opt/musa),)
-		MUSA_PATH ?= /usr/local/musa
-	else
-		MUSA_PATH ?= /opt/musa
-	endif
-	MUSA_ARCHITECTURES ?= 21;22;31
-
-	MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA
-	MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib
-	MK_LDFLAGS += -lmusa -lmusart -lmublas
-
-	ifndef GGML_NO_OPENMP
-		# For Ubuntu Focal
-		MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
-		MK_LDFLAGS  += -L/usr/lib/llvm-10/lib
-		# For Ubuntu Jammy
-		MK_CPPFLAGS += -I/usr/lib/llvm-14/lib/clang/14.0.0/include
-		MK_LDFLAGS  += -L/usr/lib/llvm-14/lib
-	endif # GGML_NO_OPENMP
-
-	CC  := $(MUSA_PATH)/bin/clang
-	CXX := $(MUSA_PATH)/bin/clang++
-	MCC := $(CCACHE) $(MUSA_PATH)/bin/mcc
-
-	MUSAFLAGS  = -fsigned-char -x musa -mtgpu
-	MUSAFLAGS += $(foreach arch,$(subst ;, ,$(MUSA_ARCHITECTURES)),--cuda-gpu-arch=mp_$(arch))
-
-ifdef GGML_CUDA_FORCE_MMQ
-	MUSAFLAGS += -DGGML_CUDA_FORCE_MMQ
-endif # GGML_CUDA_FORCE_MMQ
-
-ifdef GGML_CUDA_FORCE_CUBLAS
-	MUSAFLAGS += -DGGML_CUDA_FORCE_CUBLAS
-endif # GGML_CUDA_FORCE_CUBLAS
-
-ifdef GGML_CUDA_F16
-	MUSAFLAGS += -DGGML_CUDA_F16
-endif # GGML_CUDA_F16
-
-ifdef GGML_CUDA_DMMV_F16
-	MUSAFLAGS += -DGGML_CUDA_F16
-endif # GGML_CUDA_DMMV_F16
-
-ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
-	MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
-else
-	MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
-endif # GGML_CUDA_PEER_MAX_BATCH_SIZE
-
-ifdef GGML_CUDA_NO_PEER_COPY
-	MUSAFLAGS += -DGGML_CUDA_NO_PEER_COPY
-endif # GGML_CUDA_NO_PEER_COPY
-
-ifdef GGML_CUDA_NO_FA
-	MUSAFLAGS += -DGGML_CUDA_NO_FA
-endif # GGML_CUDA_NO_FA
-
-ifdef GGML_CUDA_FA_ALL_QUANTS
-	MUSAFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
-endif # GGML_CUDA_FA_ALL_QUANTS
-
-	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
-
-ggml/src/ggml-cuda/ggml-cuda.o: \
-	ggml/src/ggml-cuda/ggml-cuda.cu \
-	ggml/include/ggml-cuda.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h \
-	ggml/src/ggml-backend-impl.h \
-	ggml/src/ggml-common.h \
-	$(wildcard ggml/src/ggml-cuda/*.cuh)
-	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -c -o $@ $<
-
-ggml/src/ggml-cuda/%.o: \
-	ggml/src/ggml-cuda/%.cu \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h \
-	ggml/src/ggml-cuda/common.cuh
-	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -c -o $@ $<
-endif # GGML_MUSA
-
-ifdef GGML_METAL
-	MK_CPPFLAGS  += -DGGML_USE_METAL
-	MK_LDFLAGS   += -framework Foundation -framework Metal -framework MetalKit
-	OBJ_GGML_EXT += ggml/src/ggml-metal/ggml-metal.o
-
-ifdef GGML_METAL_USE_BF16
-	MK_CPPFLAGS += -DGGML_METAL_USE_BF16
-endif # GGML_METAL_USE_BF16
-ifdef GGML_METAL_NDEBUG
-	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
-endif
-ifdef GGML_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS  += -DGGML_METAL_EMBED_LIBRARY
-	OBJ_GGML_EXT += ggml/src/ggml-metal-embed.o
-endif
-endif # GGML_METAL
-
-ifdef GGML_METAL
-ggml/src/ggml-metal/ggml-metal.o: \
-	ggml/src/ggml-metal/ggml-metal.m \
-	ggml/src/ggml-metal/ggml-metal-impl.h \
-	ggml/include/ggml-metal.h \
-	ggml/include/ggml.h
-	$(CC) $(CFLAGS) -c $< -o $@
-
-ifdef GGML_METAL_EMBED_LIBRARY
-ggml/src/ggml-metal-embed.o: \
-	ggml/src/ggml-metal/ggml-metal.metal \
-	ggml/src/ggml-metal/ggml-metal-impl.h \
-	ggml/src/ggml-common.h
-	@echo "Embedding Metal library"
-	@sed -e '/__embed_ggml-common.h__/r      ggml/src/ggml-common.h'                -e '/__embed_ggml-common.h__/d'      < ggml/src/ggml-metal/ggml-metal.metal           > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp
-	@sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
-	@echo ".section __DATA, __ggml_metallib"                       >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".globl _ggml_metallib_start"                            >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo "_ggml_metallib_start:"                                  >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".globl _ggml_metallib_end"                              >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo "_ggml_metallib_end:"                                    >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
-	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
-	@rmdir ${TEMP_ASSEMBLY}
-endif
-endif # GGML_METAL
-
-DIR_GGML = ggml
-DIR_LLAMA = src
-DIR_COMMON = common
-
-OBJ_GGML = \
-	$(DIR_GGML)/src/ggml.o \
-	$(DIR_GGML)/src/ggml-alloc.o \
-	$(DIR_GGML)/src/ggml-backend.o \
-	$(DIR_GGML)/src/ggml-backend-reg.o \
-	$(DIR_GGML)/src/ggml-opt.o \
-	$(DIR_GGML)/src/ggml-quants.o \
-	$(DIR_GGML)/src/ggml-threading.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
-	$(DIR_GGML)/src/ggml-cpu/repack.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \
-	$(OBJ_GGML_EXT)
-
-OBJ_LLAMA = \
-	$(DIR_LLAMA)/llama.o \
-	$(DIR_LLAMA)/llama-vocab.o \
-	$(DIR_LLAMA)/llama-grammar.o \
-	$(DIR_LLAMA)/llama-sampling.o \
-	$(DIR_LLAMA)/unicode.o \
-	$(DIR_LLAMA)/unicode-data.o
-
-OBJ_COMMON = \
-	$(DIR_COMMON)/common.o \
-	$(DIR_COMMON)/arg.o \
-	$(DIR_COMMON)/log.o \
-	$(DIR_COMMON)/console.o \
-	$(DIR_COMMON)/ngram-cache.o \
-	$(DIR_COMMON)/sampling.o \
-	$(DIR_COMMON)/speculative.o \
-	$(DIR_COMMON)/chat.o \
-	$(DIR_COMMON)/build-info.o \
-	$(DIR_COMMON)/json-schema-to-grammar.o
-
-OBJ_ALL = $(OBJ_GGML) $(OBJ_LLAMA) $(OBJ_COMMON)
-
-LIB_GGML   = $(LIB_PRE)ggml$(DSO_EXT)
-LIB_GGML_S = $(LIB_PRE)ggml.a
-
-LIB_LLAMA   = $(LIB_PRE)llama$(DSO_EXT)
-LIB_LLAMA_S = $(LIB_PRE)llama.a
-
-LIB_COMMON   = $(LIB_PRE)common$(DSO_EXT)
-LIB_COMMON_S = $(LIB_PRE)common.a
-
-LIB_ALL   = $(LIB_GGML)   $(LIB_LLAMA)   $(LIB_COMMON)
-LIB_ALL_S = $(LIB_GGML_S) $(LIB_LLAMA_S) $(LIB_COMMON_S)
-
-GF_CC := $(CC)
-include scripts/get-flags.mk
-
-# combine build flags with cmdline overrides
-override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
-override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
-override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
-override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
-
-# identify CUDA host compiler
-ifdef GGML_CUDA
-GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
-include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
-endif
-
-ifdef LLAMA_CURL
-override CXXFLAGS := $(CXXFLAGS) -DLLAMA_USE_CURL
-override LDFLAGS  := $(LDFLAGS) -lcurl
-endif
-
-#
-# Print build information
-#
-
-$(info I llama.cpp build info: )
-$(info I UNAME_S:   $(UNAME_S))
-$(info I UNAME_P:   $(UNAME_P))
-$(info I UNAME_M:   $(UNAME_M))
-$(info I CFLAGS:    $(CFLAGS))
-$(info I CXXFLAGS:  $(CXXFLAGS))
-$(info I NVCCFLAGS: $(NVCCFLAGS))
-$(info I LDFLAGS:   $(LDFLAGS))
-$(info I CC:        $(shell $(CC)   --version | head -n 1))
-$(info I CXX:       $(shell $(CXX)  --version | head -n 1))
-ifdef GGML_CUDA
-$(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
-CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
-ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
-
-ifndef CUDA_DOCKER_ARCH
-ifndef CUDA_POWER_ARCH
-$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via environment variable CUDA_DOCKER_ARCH, e.g. by running "export CUDA_DOCKER_ARCH=compute_XX" on Unix-like systems, where XX is the minimum compute capability that the code needs to run on. A list with compute capabilities can be found here: https://developer.nvidia.com/cuda-gpus )
-endif # CUDA_POWER_ARCH
-endif # CUDA_DOCKER_ARCH
-
-endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
-endif # GGML_CUDA
-$(info )
-
-ifdef DEPRECATE_WARNING
-$(info !!! DEPRECATION WARNING !!!)
-$(info The following LLAMA_ options are deprecated and will be removed in the future. Use the GGML_ prefix instead)
-$(info   - LLAMA_CUDA)
-$(info   - LLAMA_METAL)
-$(info   - LLAMA_METAL_EMBED_LIBRARY)
-$(info   - LLAMA_OPENMP)
-$(info   - LLAMA_RPC)
-$(info   - LLAMA_SYCL)
-$(info   - LLAMA_SYCL_F16)
-$(info   - LLAMA_OPENBLAS)
-$(info   - LLAMA_OPENBLAS64)
-$(info   - LLAMA_BLIS)
-$(info   - LLAMA_NO_LLAMAFILE)
-$(info   - LLAMA_NO_ACCELERATE)
-$(info   - LLAMA_NO_OPENMP)
-$(info   - LLAMA_NO_METAL)
-$(info   - LLAMA_NO_CCACHE)
-$(info )
-endif
-
-ifdef REMOVE_WARNING
-$(info !!! REMOVAL WARNING !!!)
-$(info The following LLAMA_ options have been removed and are no longer supported)
-$(info   - LLAMA_DISABLE_LOGS   (https://github.com/ggml-org/llama.cpp/pull/9418))
-$(info   - LLAMA_SERVER_VERBOSE (https://github.com/ggml-org/llama.cpp/pull/9418))
-$(info )
-endif
-
-#
-# Build libraries
-#
-
-# Libraries
-LIB_GGML   = libggml.so
-LIB_GGML_S = libggml.a
-
-LIB_LLAMA   = libllama.so
-LIB_LLAMA_S = libllama.a
-
-LIB_COMMON   = libcommon.so
-LIB_COMMON_S = libcommon.a
-
-# Targets
-BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S)
-
-# Dependency files
-DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
-
-# Default target
-all: $(BUILD_TARGETS)
-
-# force c++ build for source file that have same name as c file
-# Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files
-$(DIR_GGML)/%_cpp.o: $(DIR_GGML)/%.cpp
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
-
-# Rules for building object files
-$(DIR_GGML)/%.o: $(DIR_GGML)/%.c
-	$(CC) $(CFLAGS) -MMD -c $< -o $@
-
-$(DIR_GGML)/%.o: $(DIR_GGML)/%.cpp
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
-
-$(DIR_LLAMA)/%.o: $(DIR_LLAMA)/%.cpp
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
-
-$(DIR_COMMON)/%.o: $(DIR_COMMON)/%.cpp
-	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
-
-# Rules for building libraries
-$(LIB_GGML): $(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
-
-$(LIB_GGML_S): $(OBJ_GGML)
-	ar rcs $(LIB_GGML_S) $^
-
-$(LIB_LLAMA): $(OBJ_LLAMA) $(LIB_GGML)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
-
-$(LIB_LLAMA_S): $(OBJ_LLAMA)
-	ar rcs $(LIB_LLAMA_S) $^
-
-$(LIB_COMMON): $(OBJ_COMMON) $(LIB_LLAMA) $(LIB_GGML)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
-
-$(LIB_COMMON_S): $(OBJ_COMMON)
-	ar rcs $(LIB_COMMON_S) $^
-
-# Include dependency files
--include $(DEP_FILES)
-
-# Clean generated server assets
-clean-server-assets:
-	find tools/server -type f -name "*.js.hpp"   -delete
-	find tools/server -type f -name "*.mjs.hpp"  -delete
-	find tools/server -type f -name "*.css.hpp"  -delete
-	find tools/server -type f -name "*.html.hpp" -delete
-
-# Clean rule
-clean: clean-server-assets
-	rm -vrf $(BUILD_TARGETS) $(TEST_TARGETS)
-	rm -rvf *.a *.dll *.so *.dot
-	find ggml src common tests examples pocs -type f -name "*.o" -delete
-	find ggml src common tests examples pocs -type f -name "*.d" -delete
-
-#
-# Examples
-#
-
-# $< is the first prerequisite, i.e. the source file.
-# Explicitly compile this to an object file so that it can be cached with ccache.
-# The source file is then filtered out from $^ (the list of all prerequisites) and the object file is added instead.
-
-# Helper function that replaces .c, .cpp, and .cu file endings with .o:
-GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))
-
-llama-cli: tools/main/main.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-	@echo
-	@echo '====  Run ./llama-cli -h for help.  ===='
-	@echo
-
-llama-run: tools/run/run.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-simple: examples/simple/simple.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-simple-chat: examples/simple-chat/simple-chat.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-tokenize: tools/tokenize/tokenize.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-batched: examples/batched/batched.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-batched-bench: tools/batched-bench/batched-bench.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-quantize: tools/quantize/quantize.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-quantize-stats: tools/quantize-stats/quantize-stats.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-perplexity: tools/perplexity/perplexity.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-imatrix: tools/imatrix/imatrix.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-embedding: examples/embedding/embedding.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-gritlm: examples/gritlm/gritlm.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-save-load-state: examples/save-load-state/save-load-state.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-gguf: examples/gguf/gguf.cpp \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-examples/gguf-hash/deps/sha1/sha1.o: \
-	examples/gguf-hash/deps/sha1/sha1.c
-	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
-
-examples/gguf-hash/deps/xxhash/xxhash.o: \
-	examples/gguf-hash/deps/xxhash/xxhash.c
-	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
-
-examples/gguf-hash/deps/sha256/sha256.o: \
-	examples/gguf-hash/deps/sha256/sha256.c
-	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
-
-llama-gguf-hash: examples/gguf-hash/gguf-hash.cpp examples/gguf-hash/deps/sha1/sha1.o examples/gguf-hash/deps/xxhash/xxhash.o examples/gguf-hash/deps/sha256/sha256.o\
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Iexamples/gguf-hash/deps -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-gguf-split: tools/gguf-split/gguf-split.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-eval-callback: examples/eval-callback/eval-callback.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-cvector-generator: tools/cvector-generator/cvector-generator.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-bench: tools/llama-bench/llama-bench.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-export-lora: tools/export-lora/export-lora.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-retrieval: examples/retrieval/retrieval.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-speculative: examples/speculative/speculative.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-parallel: examples/parallel/parallel.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-lookahead: examples/lookahead/lookahead.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-lookup: examples/lookup/lookup.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-lookup-create: examples/lookup/lookup-create.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-lookup-merge: examples/lookup/lookup-merge.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-lookup-stats: examples/lookup/lookup-stats.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-passkey: examples/passkey/passkey.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-ifdef GGML_RPC
-rpc-server: tools/rpc/rpc-server.cpp \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-endif # GGML_RPC
-
-llama-server: \
-	tools/server/server.cpp \
-	tools/server/utils.hpp \
-	tools/server/httplib.h \
-	tools/server/index.html.hpp \
-	tools/server/loading.html.hpp \
-	common/chat.cpp \
-	common/chat.h \
-	common/chat-template.hpp \
-	common/json.hpp \
-	common/minja.hpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Itools/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
-
-# Portable equivalent of `cd tools/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
-tools/server/%.hpp: tools/server/public/% FORCE Makefile
-	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
-		echo "unsigned char $${NAME}[] = {" && \
-		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
-		echo "};" && \
-		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
-	) > $@
-
-llama-gen-docs: examples/gen-docs/gen-docs.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-libllava.a: tools/mtmd/llava.cpp \
-	tools/mtmd/llava.h \
-	tools/mtmd/clip.cpp \
-	tools/mtmd/clip.h \
-	common/stb_image.h \
-	common/base64.hpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
-
-llama-llava-cli: tools/mtmd/llava-cli.cpp \
-	tools/mtmd/llava.cpp \
-	tools/mtmd/llava.h \
-	tools/mtmd/clip.cpp \
-	tools/mtmd/clip.h \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
-
-llama-minicpmv-cli: tools/mtmd/minicpmv-cli.cpp \
-	tools/mtmd/llava.cpp \
-	tools/mtmd/llava.h \
-	tools/mtmd/clip.cpp \
-	tools/mtmd/clip.h \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
-
-llama-qwen2vl-cli: tools/mtmd/qwen2vl-cli.cpp \
-	tools/mtmd/llava.cpp \
-	tools/mtmd/llava.h \
-	tools/mtmd/clip.cpp \
-	tools/mtmd/clip.h \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
-
-ifeq ($(UNAME_S),Darwin)
-swift: examples/batched.swift
-	(cd examples/batched.swift; make build)
-endif
-
-common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
-	@sh scripts/build-info.sh "$(CC)" > $@.tmp
-	@if ! cmp -s $@.tmp $@; then \
-		mv $@.tmp $@; \
-	else \
-		rm $@.tmp; \
-	fi
-
-common/build-info.o: common/build-info.cpp
-	$(CXX) $(CXXFLAGS) -c $(filter-out %.h,$^) -o $@
-
-#
-# Tests
-#
-
-tests: $(TEST_TARGETS)
-
-tests/test-arg-parser: tests/test-arg-parser.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-llama-grammar: tests/test-llama-grammar.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-log: tests/test-log.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-grammar-parser: tests/test-grammar-parser.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-grammar-integration: tests/test-grammar-integration.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-double-float: tests/test-double-float.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-chat: tests/test-chat.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-opt: tests/test-opt.cpp \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-quantize-fns: tests/test-quantize-fns.cpp \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-quantize-perf: tests/test-quantize-perf.cpp \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-sampling: tests/test-sampling.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-tokenizer-0: tests/test-tokenizer-0.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-tokenizer-1-spm: tests/test-tokenizer-1-spm.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-rope: tests/test-rope.cpp ggml/src/ggml.o \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-c.o: tests/test-c.c include/llama.h
-	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
-
-tests/test-backend-ops: tests/test-backend-ops.cpp \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-model-load-cancel: tests/test-model-load-cancel.cpp tests/get-model.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-autorelease: tests/test-autorelease.cpp tests/get-model.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-chat-template: tests/test-chat-template.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-#
-# PoCs
-#
-
-llama-vdot: pocs/vdot/vdot.cpp ggml/src/ggml.o \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-#
-# Deprecated binaries that we want to keep around long enough for people to migrate to the new filenames, then these can be removed.
-#
-# Mark legacy binary targets as .PHONY so that they are always checked.
-.PHONY: FORCE main quantize perplexity embedding server
-
-# Define the object file target
-examples/deprecation-warning/deprecation-warning.o: examples/deprecation-warning/deprecation-warning.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-# NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate.
-#  Eventually we will want to remove these target from building all the time.
-main: examples/deprecation-warning/deprecation-warning.o
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead."
-
-server: examples/deprecation-warning/deprecation-warning.o
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead."
-
-quantize: examples/deprecation-warning/deprecation-warning.o
-ifneq (,$(wildcard quantize))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "#########"
-	@echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead."
-	@echo "  Remove the 'quantize' binary to remove this warning."
-	@echo "#########"
-endif
-
-perplexity: examples/deprecation-warning/deprecation-warning.o
-ifneq (,$(wildcard perplexity))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "#########"
-	@echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead."
-	@echo "  Remove the 'perplexity' binary to remove this warning."
-	@echo "#########"
-endif
-
-embedding: examples/deprecation-warning/deprecation-warning.o
-ifneq (,$(wildcard embedding))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "#########"
-	@echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead."
-	@echo "  Remove the 'embedding' binary to remove this warning."
-	@echo "#########"
-endif
+$(error Build system changed:$(newline)\
+The Makefile build has been replaced by CMake.$(newline)$(newline)\
+For build instructions see:$(newline)\
+https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md$(newline)${newline})
diff --git a/README.md b/README.md
index 954fff83dac08..a01ef6d503e40 100644
--- a/README.md
+++ b/README.md
@@ -17,6 +17,8 @@ LLM inference in C/C++
 
 ## Hot topics
 
+- **[guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)**
+- **[[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)**
 - Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
 - Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
 - Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
@@ -149,6 +151,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 - [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
 - [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
+- [x] [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)
 
 </details>
 
@@ -240,7 +243,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 <details>
 <summary>Infrastructure</summary>
 
-- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
+- [Paddler](https://github.com/intentee/paddler) - Open-source LLMOps platform for hosting and scaling AI in your own infrastructure
 - [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 - [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
 - [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
diff --git a/ci/run.sh b/ci/run.sh
index 4d3abf9232212..d51ba443859c2 100755
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -106,7 +106,7 @@ function gg_wget {
     cd $out
 
     # should not re-download if file is the same
-    wget -nv -N $url
+    wget -nv -c -N $url
 
     cd $cwd
 }
diff --git a/common/arg.cpp b/common/arg.cpp
index 0f01bb31454a4..81c4005c5e7fc 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -749,6 +749,39 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
 // utils
 //
 
+// Helper function to parse tensor buffer override strings
+static void parse_tensor_buffer_overrides(const std::string & value, std::vector<llama_model_tensor_buft_override> & overrides) {
+    std::map<std::string, ggml_backend_buffer_type_t> buft_list;
+    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+        auto * dev = ggml_backend_dev_get(i);
+        auto * buft = ggml_backend_dev_buffer_type(dev);
+        if (buft) {
+            buft_list[ggml_backend_buft_name(buft)] = buft;
+        }
+    }
+
+    for (const auto & override : string_split<std::string>(value, ',')) {
+        std::string::size_type pos = override.find('=');
+        if (pos == std::string::npos) {
+            throw std::invalid_argument("invalid value");
+        }
+        std::string tensor_name = override.substr(0, pos);
+        std::string buffer_type = override.substr(pos + 1);
+
+        if (buft_list.find(buffer_type) == buft_list.end()) {
+            printf("Available buffer types:\n");
+            for (const auto & it : buft_list) {
+                printf("  %s\n", ggml_backend_buft_name(it.second));
+            }
+            throw std::invalid_argument("unknown buffer type");
+        }
+        // keep strings alive and avoid leaking memory by storing them in a static vector
+        static std::list<std::string> buft_overrides;
+        buft_overrides.push_back(tensor_name);
+        overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
+    }
+}
+
 struct handle_model_result {
     bool found_mmproj = false;
     common_params_model mmproj;
@@ -993,6 +1026,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
+    if (!params.speculative.tensor_buft_overrides.empty()) {
+        params.speculative.tensor_buft_overrides.push_back({nullptr, nullptr});
+    }
+
     if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
         throw std::runtime_error(string_format(
             "error: the supplied chat template is not supported: %s%s\n",
@@ -1201,6 +1238,7 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
             common_params_print_completion(ctx_arg);
             exit(0);
         }
+        params.lr.init();
     } catch (const std::invalid_argument & ex) {
         fprintf(stderr, "%s\n", ex.what());
         ctx_arg.params = params_org;
@@ -1469,6 +1507,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.swa_full = true;
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
+    add_opt(common_arg(
+        {"--swa-checkpoints"}, "N",
+        string_format("max number of SWA checkpoints per slot to create (default: %d)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_swa_checkpoints),
+        [](common_params & params, int value) {
+            params.n_swa_checkpoints = value;
+        }
+    ).set_env("LLAMA_ARG_SWA_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--kv-unified", "-kvu"},
         string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
@@ -1484,6 +1530,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.ctx_shift = false;
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
+    add_opt(common_arg(
+        {"--context-shift"},
+        string_format("enables context shift on infinite text generation (default: %s)", params.ctx_shift ? "enabled" : "disabled"),
+        [](common_params & params) {
+            params.ctx_shift = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY}).set_env("LLAMA_ARG_CONTEXT_SHIFT"));
     add_opt(common_arg(
         {"--chunks"}, "N",
         string_format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks),
@@ -1702,7 +1755,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -1777,7 +1830,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.sampling.top_n_sigma = std::stof(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}).set_sparam());
+    ).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
         string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
@@ -2201,9 +2254,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"-dt", "--defrag-thold"}, "N",
-        string_format("KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold),
+        string_format("KV cache defragmentation threshold (DEPRECATED)"),
         [](common_params & params, const std::string & value) {
-            params.defrag_thold = std::stof(value);
+            GGML_UNUSED(params);
+            GGML_UNUSED(value);
+            LOG_WRN("DEPRECATED: --defrag-thold is deprecated and no longer necessary to specify\n");
         }
     ).set_env("LLAMA_ARG_DEFRAG_THOLD"));
     add_opt(common_arg(
@@ -2349,40 +2404,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--override-tensor", "-ot"}, "<tensor name pattern>=<buffer type>,...",
         "override tensor buffer type", [](common_params & params, const std::string & value) {
-            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
-            if (buft_list.empty()) {
-                // enumerate all the devices and add their buffer types to the list
-                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-                    auto * dev = ggml_backend_dev_get(i);
-                    auto * buft = ggml_backend_dev_buffer_type(dev);
-                    if (buft) {
-                        buft_list[ggml_backend_buft_name(buft)] = buft;
-                    }
-                }
-            }
-
-            for (const auto & override : string_split<std::string>(value, ',')) {
-                std::string::size_type pos = override.find('=');
-                if (pos == std::string::npos) {
-                    throw std::invalid_argument("invalid value");
-                }
-                std::string tensor_name = override.substr(0, pos);
-                std::string buffer_type = override.substr(pos + 1);
-
-                if (buft_list.find(buffer_type) == buft_list.end()) {
-                    printf("Available buffer types:\n");
-                    for (const auto & it : buft_list) {
-                        printf("  %s\n", ggml_backend_buft_name(it.second));
-                    }
-                    throw std::invalid_argument("unknown buffer type");
-                }
-                // keep strings alive and avoid leaking memory by storing them in a static vector
-                static std::list<std::string> buft_overrides;
-                buft_overrides.push_back(tensor_name);
-                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
-            }
+            parse_tensor_buffer_overrides(value, params.tensor_buft_overrides);
         }
     ));
+    add_opt(common_arg(
+        {"--override-tensor-draft", "-otd"}, "<tensor name pattern>=<buffer type>,...",
+        "override tensor buffer type for draft model", [](common_params & params, const std::string & value) {
+            parse_tensor_buffer_overrides(value, params.speculative.tensor_buft_overrides);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--cpu-moe", "-cmoe"},
         "keep all Mixture of Experts (MoE) weights in the CPU",
@@ -2405,6 +2435,27 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_N_CPU_MOE"));
+    add_opt(common_arg(
+        {"--cpu-moe-draft", "-cmoed"},
+        "keep all Mixture of Experts (MoE) weights in the CPU for the draft model",
+        [](common_params & params) {
+            params.speculative.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CPU_MOE_DRAFT"));
+    add_opt(common_arg(
+        {"--n-cpu-moe-draft", "-ncmoed"}, "N",
+        "keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model",
+        [](common_params & params, int value) {
+            if (value < 0) {
+                throw std::invalid_argument("invalid value");
+            }
+            for (int i = 0; i < value; ++i) {
+                static std::list<std::string> buft_overrides_draft;
+                buft_overrides_draft.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                params.speculative.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_CPU_MOE_DRAFT"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -2655,7 +2706,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.out_file = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS}));
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_FINETUNE}));
     add_opt(common_arg(
         {"-ofreq", "--output-frequency"}, "N",
         string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@@ -2949,11 +3000,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)\n"
         "(default: auto)",
         [](common_params & params, const std::string & value) {
-            /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
-            else if (value == "deepseek-legacy") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY; }
-            else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
-            else if (value == "auto") {     params.reasoning_format = COMMON_REASONING_FORMAT_AUTO; }
-            else { throw std::invalid_argument("invalid value"); }
+            params.reasoning_format = common_reasoning_format_from_name(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));
     add_opt(common_arg(
@@ -3134,7 +3181,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.speculative.cpuparams.n_threads = std::thread::hardware_concurrency();
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-tbd", "--threads-batch-draft"}, "N",
         "number of threads to use during batch and prompt processing (default: same as --threads-draft)",
@@ -3144,7 +3191,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.speculative.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-Cd", "--cpu-mask-draft"}, "M",
         "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
@@ -3537,5 +3584,51 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
 
 
+    add_opt(
+        common_arg({ "-lr", "--learning-rate" }, "ALPHA",
+                   string_format(
+                       "adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)",
+                       (double) params.lr.lr0),
+                   [](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(
+        common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
+                   string_format(
+                       "(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
+                       (double) params.lr.lr_min),
+                   [](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(
+        common_arg({ "-decay-epochs", "--learning-rate-decay-epochs" }, "ALPHA",
+                   string_format(
+                       "(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)",
+                       (double) params.lr.decay_epochs),
+                   [](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+                { "-wd", "--weight-decay" }, "WD",
+                string_format(
+                    "adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).",
+                    (double) params.lr.wd),
+                [](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-val-split", "--val-split" }, "FRACTION",
+                       string_format("fraction of data to use as validation set for training (default: %.2g).",
+                                     (double) params.val_split),
+                       [](common_params & params, const std::string & value) { params.val_split = std::stof(value); })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-epochs", "--epochs" }, "N",
+                       string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
+                       [](common_params & params, int epochs) { params.lr.epochs = epochs; })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-opt", "--optimizer" }, "sgd|adamw", "adamw or sgd",
+                       [](common_params & params, const std::string & name) {
+                           params.optimizer = common_opt_get_optimizer(name.c_str());
+                           if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
+                               throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
+                           }
+                       })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+
     return ctx_arg;
 }
diff --git a/common/chat.cpp b/common/chat.cpp
index 316bd24170c9e..111b4a21b368c 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -147,6 +147,7 @@ struct templates_params {
     json extra_context;
     bool add_bos;
     bool add_eos;
+    bool is_inference = true;
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -296,6 +297,7 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
         }
         if (!msg.reasoning_content.empty()) {
             jmsg["reasoning_content"] = msg.reasoning_content;
+            jmsg["thinking"] = msg.reasoning_content; // gpt-oss
         }
         if (!msg.tool_name.empty()) {
             jmsg["name"] = msg.tool_name;
@@ -472,11 +474,12 @@ std::string common_chat_format_single(
     return ss.str();
 }
 
-std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja) {
+std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja, const std::map<std::string, std::string> & chat_template_kwargs) {
     common_chat_templates_inputs inputs;
     inputs.use_jinja = use_jinja;
     inputs.add_bos = tmpls->add_bos;
     inputs.add_eos = tmpls->add_eos;
+    inputs.chat_template_kwargs = chat_template_kwargs;
     auto add_simple_msg = [&](auto role, auto content) {
         common_chat_msg msg;
         msg.role = role;
@@ -552,6 +555,17 @@ common_chat_templates_ptr common_chat_templates_init(
             default_template_src = CHATML_TEMPLATE_SRC;
         }
     }
+
+    // TODO @ngxson : this is a temporary hack to prevent chat template from throwing an error
+    // Ref: https://github.com/ggml-org/llama.cpp/pull/15230#issuecomment-3173959633
+    if (default_template_src.find("<|channel|>") != std::string::npos
+            // search for the error message and patch it
+            && default_template_src.find("in message.content or") != std::string::npos) {
+        string_replace_all(default_template_src,
+            "{%- if \"<|channel|>analysis<|message|>\" in message.content or \"<|channel|>final<|message|>\" in message.content %}",
+            "{%- if false %}");
+    }
+
     std::string token_bos = bos_token_override;
     std::string token_eos = eos_token_override;
     bool add_bos = false;
@@ -619,12 +633,24 @@ const char * common_reasoning_format_name(common_reasoning_format format) {
         case COMMON_REASONING_FORMAT_AUTO:     return "auto";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
         case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
-        case COMMON_REASONING_FORMAT_GRANITE: return "granite";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }
 }
 
+common_reasoning_format common_reasoning_format_from_name(const std::string & format) {
+    if (format == "none") {
+        return COMMON_REASONING_FORMAT_NONE;
+    } else if (format == "auto") {
+        return COMMON_REASONING_FORMAT_AUTO;
+    } else if (format == "deepseek") {
+        return COMMON_REASONING_FORMAT_DEEPSEEK;
+    } else if (format == "deepseek-legacy") {
+        return COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY;
+    }
+    throw std::runtime_error("Unknown reasoning format: " + format);
+}
+
 static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) {
     std::string arguments;
     if (builder.is_partial()) {
@@ -1311,19 +1337,198 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
     common_chat_params data;
     auto prompt = apply(tmpl, inputs);
 
+    // Check if we need to replace the return token with end token during
+    // inference and without generation prompt. For more details see:
+    // https://github.com/ggml-org/llama.cpp/issues/15417
+    if (inputs.is_inference && !inputs.add_generation_prompt) {
+        static constexpr std::string_view return_token = "<|return|>";
+        static constexpr std::string_view end_token    = "<|end|>";
+        if (size_t pos = prompt.rfind(return_token); pos != std::string::npos) {
+            prompt.replace(pos, return_token.length(), end_token);
+        }
+    }
+
     data.prompt = prompt;
     data.format = COMMON_CHAT_FORMAT_GPT_OSS;
 
-    // TODO: support tool calls in GPT-OSS?
+    // These special tokens are required to parse properly, so we include them
+    // even if parse_tool_calls is false.
+    data.preserved_tokens = {
+        "<|channel|>",
+        "<|constrain|>",
+        "<|message|>",
+        "<|start|>",
+        "<|end|>",
+    };
+
+    if (!inputs.json_schema.is_null()) {
+        data.grammar_lazy = false;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            auto schema = inputs.json_schema;
+            builder.resolve_refs(schema);
+
+            auto not_end = builder.add_rule("not-end",
+                "[^<] | \"<\" [^|] | \"<|\" [^e] | \"<|e\" [^n] | \"<|en\" [^d] | \"<|end\" [^|] | \"<|end|\" [^>]");
+            auto analysis = builder.add_rule("analysis",
+                "\"<|channel|>analysis<|message|>\" ( " + not_end + " )* \"<|end|>\"");
+            auto constraint = builder.add_rule("constraint", "\"<|constrain|>\"? [a-zA-Z0-9_-]+");
+            auto final = builder.add_rule("final",
+                "\"<|channel|>final\" ( \" \" " + constraint + " )? \"<|message|>\" " +
+                builder.add_schema("response", schema)
+            );
+
+            builder.add_rule("root", "( " + analysis + " \"<|start|>assistant\" )? " + final);
+        });
+    }
+
+    if (inputs.tools.is_array() && !inputs.tools.empty()) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            // tool calls can appear in commentary or analysis channels
+            auto channel = builder.add_rule("channel", "\"<|channel|>\" ( \"commentary\" | \"analysis\" )");
+
+            std::vector<std::string> tool_rules_recipient_in_role;
+            std::vector<std::string> tool_rules_recipient_in_channel;
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+                builder.resolve_refs(parameters);
+
+                tool_rules_recipient_in_role.push_back(
+                    builder.add_rule(name + "-call",
+                        "\"" + name + "\"" + channel + " \" <|constrain|>json\"? \"<|message|>\" " +
+                        builder.add_schema(name + "-args", parameters)
+                    )
+                );
+
+                tool_rules_recipient_in_channel.push_back(
+                    builder.add_rule(name + "-call",
+                        "\"" + name + "\"" + " \" <|constrain|>json\"? \"<|message|>\" " +
+                        builder.add_schema(name + "-args", parameters)
+                    )
+                );
+            });
+
+            auto recipient_in_role = builder.add_rule("recipient_in_role",
+                "\"<|start|>assistant\"? \" to=functions.\" ( " +
+                string_join(tool_rules_recipient_in_role, " | ") + " )"
+            );
+
+            auto recipient_in_channel = builder.add_rule("recipient_in_channel",
+                channel + " \" to=functions.\" ( " +
+                string_join(tool_rules_recipient_in_channel, " | ") + " )"
+            );
+
+            builder.add_rule("root", recipient_in_role + " | " + recipient_in_channel);
+
+            // Trigger on tool calls that appear in the commentary channel
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+                "<\\|channel\\|>(commentary|analysis) to"
+            });
+
+            // Trigger tool calls that appear in the role section, either at the
+            // start or in the middle.
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
+                "^ to"
+            });
+
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+                "<\\|start\\|>assistant to"
+            });
+        });
+    }
 
     return data;
 }
 static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
-    // TODO @ngxson : this won't work with --special enabled, we should fix that
-    builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|start|>assistant<|channel|>final<|message|>");
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
+    static const std::string constraint = "(?: (<\\|constrain\\|>)?([a-zA-Z0-9_-]+))";
+    static const std::string recipient("(?: to=functions\\.([^<\\s]+))");
+
+    static const common_regex start_regex("<\\|start\\|>assistant");
+    static const common_regex analysis_regex("<\\|channel\\|>analysis");
+    static const common_regex final_regex("<\\|channel\\|>final" + constraint + "?");
+    static const common_regex preamble_regex("<\\|channel\\|>commentary");
+    static const common_regex tool_call1_regex(recipient + "<\\|channel\\|>(analysis|commentary)" + constraint + "?");
+    static const common_regex tool_call2_regex("<\\|channel\\|>(analysis|commentary)" + recipient + constraint + "?");
+
+    auto consume_end = [&](bool include_end = false) {
+        if (auto res = builder.try_find_literal("<|end|>")) {
+            return res->prelude + (include_end ? builder.str(res->groups[0]) : "");
+        }
+        return builder.consume_rest();
+    };
+
+    auto handle_tool_call = [&](const std::string & name) {
+        if (auto args = builder.try_consume_json_with_dumped_args({{}})) {
+            if (builder.syntax().parse_tool_calls) {
+                if (!builder.add_tool_call(name, "", args->value) || args->is_partial) {
+                    throw common_chat_msg_partial_exception("incomplete tool call");
+                }
+            } else if (args->is_partial) {
+                throw common_chat_msg_partial_exception("incomplete tool call");
+            }
+        }
+    };
+
+    auto regex_match = [](const common_regex & regex, const std::string & input) -> std::optional<common_regex_match> {
+        auto match = regex.search(input, 0, true);
+        if (match.type == COMMON_REGEX_MATCH_TYPE_FULL) {
+            return match;
+        }
+        return std::nullopt;
+    };
+
+    do {
+        auto header_start_pos = builder.pos();
+        auto content_start = builder.try_find_literal("<|message|>");
+        if (!content_start) {
+            throw common_chat_msg_partial_exception("incomplete header");
+        }
+
+        auto header = content_start->prelude;
+
+        if (auto match = regex_match(tool_call1_regex, header)) {
+            auto group = match->groups[1];
+            auto name = header.substr(group.begin, group.end - group.begin);
+            handle_tool_call(name);
+            continue;
+        }
+
+        if (auto match = regex_match(tool_call2_regex, header)) {
+            auto group = match->groups[2];
+            auto name = header.substr(group.begin, group.end - group.begin);
+            handle_tool_call(name);
+            continue;
+        }
+
+        if (regex_match(analysis_regex, header)) {
+            builder.move_to(header_start_pos);
+            if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) {
+                builder.add_content(consume_end(true));
+            } else {
+                builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|end|>");
+            }
+            continue;
+        }
+
+        if(regex_match(final_regex, header) || regex_match(preamble_regex, header)) {
+            builder.add_content(consume_end());
+            continue;
+        }
+
+        // Possibly a malformed message, attempt to recover by rolling
+        // back to pick up the next <|start|>
+        LOG_DBG("%s: unknown header from message: %s\n", __func__, header.c_str());
+        builder.move_to(header_start_pos);
+    } while (builder.try_find_regex(start_regex, std::string::npos, false));
+
+    auto remaining = builder.consume_rest();
+    if (!remaining.empty()) {
+        LOG_DBG("%s: content after last message: %s\n", __func__, remaining.c_str());
     }
 }
 
@@ -1887,8 +2092,8 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
-    params.add_bos = inputs.add_bos;
-    params.add_eos = inputs.add_eos;
+    params.add_bos = tmpls->add_bos;
+    params.add_eos = tmpls->add_eos;
 
     params.extra_context = json::object();
     for (auto el : inputs.chat_template_kwargs) {
@@ -1936,7 +2141,7 @@ static common_chat_params common_chat_templates_apply_jinja(
     }
 
     // GPT-OSS
-    if (src.find("<|channel|>") != std::string::npos && params.json_schema.is_null()) {
+    if (src.find("<|channel|>") != std::string::npos) {
         return common_chat_params_init_gpt_oss(tmpl, params);
     }
 
diff --git a/common/chat.h b/common/chat.h
index eb628d8bc275d..d1e480c918794 100644
--- a/common/chat.h
+++ b/common/chat.h
@@ -187,10 +187,12 @@ std::string common_chat_format_single(
 // Returns an example of formatted chat
 std::string common_chat_format_example(
     const struct common_chat_templates * tmpls,
-    bool use_jinja);
+    bool use_jinja,
+    const std::map<std::string, std::string> & chat_template_kwargs);
 
 const char*               common_chat_format_name(common_chat_format format);
 const char*               common_reasoning_format_name(common_reasoning_format format);
+common_reasoning_format   common_reasoning_format_from_name(const std::string & format);
 common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
diff --git a/common/common.cpp b/common/common.cpp
index c6962d1d19b33..fdce1dcdec19b 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -41,6 +41,7 @@
 #endif
 #include <locale>
 #include <windows.h>
+#include <string.h>
 #include <fcntl.h>
 #include <io.h>
 #else
@@ -557,13 +558,6 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam
 
         auto detokenized = common_token_to_piece(ctx, token);
 
-        detokenized.erase(
-            std::remove_if(
-                detokenized.begin(),
-                detokenized.end(),
-                [](const unsigned char c) { return !std::isprint(c); }),
-            detokenized.end());
-
         buf << "'" << detokenized << "'"
             << ":" << std::to_string(token);
     }
@@ -588,13 +582,6 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 
         auto detokenized = common_token_to_piece(ctx, batch.token[i]);
 
-        detokenized.erase(
-                std::remove_if(
-                    detokenized.begin(),
-                    detokenized.end(),
-                    [](const unsigned char c) { return !std::isprint(c); }),
-                detokenized.end());
-
         buf << "\n"          << std::to_string(i)
             << ", token '"   << detokenized << "'"
             << ", pos "      << std::to_string(batch.pos[i])
@@ -1165,7 +1152,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
     cparams.attention_type    = params.attention_type;
-    cparams.defrag_thold      = params.defrag_thold;
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
@@ -1565,3 +1551,56 @@ ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std
 
     return result;
 }
+
+ggml_opt_optimizer_params common_opt_lr_pars(void * userdata) {
+    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
+    const lr_opt &            d      = *(lr_opt *) userdata;
+    result.adamw.alpha = result.sgd.alpha = d.get_lr(d.epoch);
+    result.sgd.wd = result.adamw.wd = d.wd;
+    return result;
+}
+
+// TODO make all command line args case-insensitive
+static inline bool eq_case_insensitive(char const* a, char const* b) {
+    return !
+#if defined(_MSC_VER)
+        _stricmp
+#else
+        strcasecmp
+#endif // defined(_MSC_VER)
+        (a, b);
+}
+
+enum ggml_opt_optimizer_type common_opt_get_optimizer(const char * n) {
+    if (eq_case_insensitive("adamw", n)) {
+        return GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    }
+    if (eq_case_insensitive("sgd", n)) {
+        return GGML_OPT_OPTIMIZER_TYPE_SGD;
+    }
+    return GGML_OPT_OPTIMIZER_TYPE_COUNT;
+}
+
+// TODO simplify to use just log and exp
+static float const k_log_2 = std::log(2.f);
+
+void lr_opt::init() {
+    if (lr_min > 0 && lr_min < lr0) {
+        float nhalf = std::log(lr0 / lr_min) / k_log_2;
+        float e     = epochs;
+        if (decay_epochs > 0 && decay_epochs < e) {
+            e = decay_epochs;
+        } else {
+            decay_epochs = e;
+        }
+        scale_epoch = nhalf / e;
+    }
+}
+
+float lr_opt::get_lr(float epoch) const {
+    float r = lr_min <= 0 ? lr0 :
+        epoch >= decay_epochs ? lr_min :
+        lr0 * std::pow(0.5f, epoch * scale_epoch);
+    LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
+    return r;
+}
diff --git a/common/common.h b/common/common.h
index 5eab199af559e..390dda5e531be 100644
--- a/common/common.h
+++ b/common/common.h
@@ -2,14 +2,17 @@
 
 #pragma once
 
-#include "llama-cpp.h"
-
 #include <set>
+#include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
 #include <sstream>
+#include <cmath>
+
+#include "ggml-opt.h"
+#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -82,6 +85,7 @@ enum llama_example {
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
+    LLAMA_EXAMPLE_FINETUNE,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -202,6 +206,7 @@ struct common_params_speculative {
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
     std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
@@ -234,14 +239,36 @@ struct common_params_diffusion {
     bool    add_gumbel_noise = false; // add gumbel noise to the logits if temp > 0.0
 };
 
+// reasoning API response format (not to be confused as chat template's reasoning format)
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_AUTO,
+    COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`
     COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
     COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
-    COMMON_REASONING_FORMAT_GRANITE,         // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
+    // do not extend this enum unless you absolutely have to
+    // in most cases, use COMMON_REASONING_FORMAT_AUTO
+    // see: https://github.com/ggml-org/llama.cpp/pull/15408
 };
 
+
+struct lr_opt {
+    float    lr0          = 1e-5; // learning rate at first epoch
+    float    lr_min       = -1;
+    float    decay_epochs = -1;   // if >0, the learning rate starts at lr0 and decays to lr_min after this many epochs
+    float    scale_epoch  = 0;
+    float    wd           = 0;
+    unsigned epochs       = 2;
+
+    unsigned epoch; // set by optimizer outer (epochs) loop
+    // learning rate decay - constant LR per epoch only for now
+    float get_lr(float e) const;
+    float get_lr() const { return get_lr(epoch); }
+    // must call after arg parse, before get_lr
+    void init();
+};
+
+struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
+
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
@@ -261,7 +288,6 @@ struct common_params {
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
     float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
-    float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
@@ -348,7 +374,7 @@ struct common_params {
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool flash_attn        = false; // flash attention
     bool no_perf           = false; // disable performance metrics
-    bool ctx_shift         = true;  // context shift on inifinite text generation
+    bool ctx_shift         = false;  // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
 
@@ -376,6 +402,11 @@ struct common_params {
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 
+    // finetune
+    struct lr_opt lr;
+    enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    float val_split = 0.05f; // fraction of the data used for the validation set
+
     // embedding
     bool embedding         = false; // get only sentence embedding
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
@@ -384,11 +415,12 @@ struct common_params {
     std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
-    int32_t port           = 8080;         // server listens on this network port
-    int32_t timeout_read   = 600;          // http read timeout in seconds
-    int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
-    int32_t n_cache_reuse  = 0;            // min chunk size to reuse from the cache via KV shifting
+    int32_t port              = 8080;         // server listens on this network port
+    int32_t timeout_read      = 600;          // http read timeout in seconds
+    int32_t timeout_write     = timeout_read; // http write timeout in seconds
+    int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
+    int32_t n_swa_checkpoints = 3;            // max number of SWA checkpoints per slot
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -703,3 +735,6 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 //
 
 ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
+
+// "adamw" or "sgd" (case insensitive)
+enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index b8c7d97a786c7..35fadbc83ea1b 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -28,6 +28,14 @@
 if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
+from gguf.vocab import MistralTokenizerType, MistralVocab
+from mistral_common.tokens.tokenizers.base import TokenizerVersion
+from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN, DATASET_STD
+from mistral_common.tokens.tokenizers.tekken import Tekkenizer
+from mistral_common.tokens.tokenizers.sentencepiece import (
+    SentencePieceTokenizer,
+)
+
 
 logger = logging.getLogger("hf-to-gguf")
 
@@ -81,11 +89,16 @@ class ModelBase:
     block_count: int
     tensor_map: gguf.TensorNameMap
 
+    # Mistral format specifics
+    is_mistral_format: bool = False
+    disable_mistral_community_chat_template: bool = False
+
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
-                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None):
+                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
+                 disable_mistral_community_chat_template: bool = False):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
                 type(self) is MmprojModel:
@@ -106,16 +119,17 @@ def get_remote_tensors() -> Iterator[tuple[str, Tensor]]:
                 logger.info(f"Using remote model with HuggingFace id: {remote_hf_model_id}")
                 remote_tensors = gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id)
                 self.tensor_names = set(name for name in remote_tensors.keys())
-                for name, remote_tensor in gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id).items():
+                for name, remote_tensor in remote_tensors.items():
                     yield (name, LazyTorchTensor.from_remote_tensor(remote_tensor))
 
             self.get_tensors = get_remote_tensors
         else:
-            self.part_names = ModelBase.get_model_part_names(self.dir_model, "model", ".safetensors")
+            prefix = "model" if not self.is_mistral_format else "consolidated"
+            self.part_names = ModelBase.get_model_part_names(self.dir_model, prefix, ".safetensors")
             self.is_safetensors = len(self.part_names) > 0
             if not self.is_safetensors:
                 self.part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
-        self.hparams = ModelBase.load_hparams(self.dir_model) if hparams is None else hparams
+        self.hparams = ModelBase.load_hparams(self.dir_model, self.is_mistral_format) if hparams is None else hparams
         self.tensor_names = None
         self.metadata_override = metadata_override
         self.model_name = model_name
@@ -136,6 +150,9 @@ def get_remote_tensors() -> Iterator[tuple[str, Tensor]]:
         self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file,
                                            split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard)
 
+        # Mistral specific
+        self.disable_mistral_community_chat_template = disable_mistral_community_chat_template
+
     @classmethod
     def add_prefix_to_filename(cls, path: Path, prefix: str) -> Path:
         stem, suffix = path.stem, path.suffix
@@ -153,19 +170,23 @@ def find_hparam(self, keys: Iterable[str], optional: bool = False) -> Any:
     def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
         tensor_names_from_parts: set[str] = set()
 
-        index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
-        index_name += ".index.json"
-        index_file = self.dir_model / index_name
-
-        if index_file.is_file():
-            self.tensor_names = set()
-            logger.info(f"gguf: loading model weight map from '{index_name}'")
-            with open(index_file, "r", encoding="utf-8") as f:
-                index: dict[str, Any] = json.load(f)
-                weight_map = index.get("weight_map")
-                if weight_map is None or not isinstance(weight_map, dict):
-                    raise ValueError(f"Can't load 'weight_map' from {index_name!r}")
-                self.tensor_names.update(weight_map.keys())
+        if not self.is_mistral_format:
+            index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
+            index_name += ".index.json"
+            index_file = self.dir_model / index_name
+
+            if index_file.is_file():
+                self.tensor_names = set()
+                logger.info(f"gguf: loading model weight map from '{index_name}'")
+                with open(index_file, "r", encoding="utf-8") as f:
+                    index: dict[str, Any] = json.load(f)
+                    weight_map = index.get("weight_map")
+                    if weight_map is None or not isinstance(weight_map, dict):
+                        raise ValueError(f"Can't load 'weight_map' from {index_name!r}")
+                    self.tensor_names.update(weight_map.keys())
+            else:
+                self.tensor_names = tensor_names_from_parts
+                weight_map = {}
         else:
             self.tensor_names = tensor_names_from_parts
             weight_map = {}
@@ -426,7 +447,12 @@ def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]
         return part_names
 
     @staticmethod
-    def load_hparams(dir_model: Path):
+    def load_hparams(dir_model: Path, is_mistral_format: bool):
+        if is_mistral_format:
+            with open(dir_model / "params.json", "r", encoding="utf-8") as f:
+                config = json.load(f)
+            return config
+
         try:
             # for security reason, we don't allow loading remote code by default
             # if a model need remote code, we will fallback to config.json
@@ -476,7 +502,10 @@ class TextModel(ModelBase):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.hf_arch = get_model_architecture(self.hparams, self.model_type)
+        if not self.is_mistral_format:
+            self.hf_arch = get_model_architecture(self.hparams, self.model_type)
+        else:
+            self.hf_arch = ""
 
         if "text_config" in self.hparams:
             # move the text_config to the root level
@@ -542,14 +571,14 @@ def set_gguf_parameters(self):
             self.gguf_writer.add_head_count(n_head)
             logger.info(f"gguf: head count = {n_head}")
 
-        if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
+        if (n_head_kv := self.find_hparam(["num_key_value_heads", "n_kv_heads"], optional=True)) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
             logger.info(f"gguf: key-value head count = {n_head_kv}")
 
         if (rope_theta := self.hparams.get("rope_theta")) is not None:
             self.gguf_writer.add_rope_freq_base(rope_theta)
             logger.info(f"gguf: rope theta = {rope_theta}")
-        if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+        if (f_rms_eps := self.find_hparam(["rms_norm_eps", "norm_eps"], optional=True)) is not None:
             self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
             logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
         if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
@@ -1210,12 +1239,19 @@ def __init__(self, *args, **kwargs):
             raise TypeError("MmprojModel must be subclassed with model_arch = gguf.MODEL_ARCH.MMPROJ")
 
         # get n_embd of the text model
-        if "text_config" not in self.hparams:
-            self.hparams["text_config"] = {}
-        if "audio_config" not in self.hparams:
-            self.hparams["audio_config"] = {}
-        text_config = {**self.hparams, **self.hparams["text_config"]}
-        self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
+        if not self.is_mistral_format:
+            if "text_config" not in self.hparams:
+                self.hparams["text_config"] = {}
+            if "audio_config" not in self.hparams:
+                self.hparams["audio_config"] = {}
+            text_config = {**self.hparams, **self.hparams["text_config"]}
+            self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
+        else:
+            text_config = {
+                k: v for k, v in self.hparams.items() if k not in ["vision_encoder", "audio_encoder"]
+            }
+            self.n_embd_text = text_config.get("hidden_dim", 0)
+
         assert self.n_embd_text > 0, "n_embd not found in hparams"
 
         # move vision config to the top level, while preserving the original hparams in global_config
@@ -1236,11 +1272,13 @@ def __init__(self, *args, **kwargs):
         self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.MMPROJ, self.block_count)
 
         # load preprocessor config
-        with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
-            self.preprocessor_config = json.load(f)
+        if not self.is_mistral_format:
+            with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
+                self.preprocessor_config = json.load(f)
 
     def get_vision_config(self) -> dict[str, Any] | None:
-        return self.global_config.get("vision_config")
+        config_name = "vision_config" if not self.is_mistral_format else "vision_encoder"
+        return self.global_config.get(config_name)
 
     def get_audio_config(self) -> dict[str, Any] | None:
         return self.global_config.get("audio_config")
@@ -1264,8 +1302,11 @@ def set_gguf_parameters(self):
             self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads"]))
 
             # preprocessor config
-            self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
-            self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
+            image_mean = DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]
+            image_std = DATASET_STD if self.is_mistral_format else self.preprocessor_config["image_std"]
+
+            self.gguf_writer.add_vision_image_mean(image_mean)
+            self.gguf_writer.add_vision_image_std(image_std)
 
         if self.has_audio_encoder:
             self.gguf_writer.add_clip_has_audio_encoder(True)
@@ -1299,6 +1340,12 @@ def _find_param(self, obj: dict[str, Any], keys: Iterable[str], optional: bool =
             return None
         raise KeyError(f"could not find any of: {keys}")
 
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, name, n_dims  # unused
+        if ".patch_embd.weight" in new_name:
+            return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
+        return False
+
 
 @ModelBase.register("GPTNeoXForCausalLM")
 class GPTNeoXModel(TextModel):
@@ -1924,46 +1971,12 @@ def __init__(self, *args, **kwargs):
         if self.hf_arch == "VLlama3ForCausalLM":
             self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
 
-    def set_vocab(self):
-        path_tekken_json = self.dir_model / "tekken.json"
-        path_tokenizer_json = self.dir_model / "tokenizer.json"
-        if path_tekken_json.is_file() and not path_tokenizer_json.is_file():
-            return self.set_vocab_tekken()
-
-        try:
-            self._set_vocab_sentencepiece()
-        except FileNotFoundError:
-            try:
-                self._set_vocab_llama_hf()
-            except (FileNotFoundError, TypeError):
-                # Llama 3
-                self._set_vocab_gpt2()
-
-        # Apply to CodeLlama only (and ignore for Llama 3 with a vocab size of 128256)
-        if self.hparams.get("vocab_size", 32000) == 32016:
-            special_vocab = gguf.SpecialVocab(
-                self.dir_model, load_merges=False,
-                special_token_types = ['prefix', 'suffix', 'middle', 'eot']
-            )
-            special_vocab._set_special_token("prefix", 32007)
-            special_vocab._set_special_token("suffix", 32008)
-            special_vocab._set_special_token("middle", 32009)
-            special_vocab._set_special_token("eot",    32010)
-            special_vocab.add_to_gguf(self.gguf_writer)
-
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                if "add_prefix_space" in tokenizer_config_json:
-                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
-
-        # Apply to granite small models only
-        if self.hparams.get("vocab_size", 32000) == 49152:
-            self.gguf_writer.add_add_bos_token(False)
+    def _set_vocab_mistral(self):
+        vocab = MistralVocab(self.dir_model)
+        logger.info(
+            f"Converting tokenizer {vocab.tokenizer_type} of size {vocab.vocab_size}."
+        )
 
-    def set_vocab_tekken(self):
-        vocab = gguf.vocab.MistralVocab(self.dir_model)
         self.gguf_writer.add_tokenizer_model(vocab.gguf_tokenizer_model)
 
         tokens = []
@@ -1979,7 +1992,7 @@ def set_vocab_tekken(self):
             f"token count ({len(tokens)}) != vocab size ({vocab.vocab_size})"
         )
 
-        if vocab.tokenizer_type == gguf.vocab.MistralTokenizerType.tekken:
+        if vocab.tokenizer_type == MistralTokenizerType.tekken:
             self.gguf_writer.add_tokenizer_pre("tekken")
             self.gguf_writer.add_token_merges(
                 vocab.extract_vocab_merges_from_model()
@@ -2002,16 +2015,67 @@ def set_vocab_tekken(self):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(False)
 
-        script_dir = Path(__file__).parent
-        template_path = script_dir / "models/templates/unsloth-mistral-Devstral-Small-2507.jinja"
-        with open(template_path, "r", encoding="utf-8") as f:
-            template = f.read()
+        template_dir = Path(__file__).parent / "models/templates/"
+
+        if not self.is_mistral_format or not self.disable_mistral_community_chat_template:
+            # Log only for Mistral format that the official tokenization and detokenization is via `mistral-common`.
+            if self.is_mistral_format:
+                logger.info(
+                    "Using a Mistral community chat template. These templates can be subject to errors in early days or weeks after a release. "
+                    "Mistral recommends to use `mistral-common` to perform tokenization and detokenization."
+                )
+            template = MistralModel.get_community_chat_template(vocab, template_dir, self.is_mistral_format)
             self.gguf_writer.add_chat_template(template)
+        else:
+            logger.info("Not using a Mistral community chat template. Ensure to perform the tokenization and detokenization via `mistral-common`.")
+
+    def set_vocab(self):
+        if self.is_mistral_format:
+            return self._set_vocab_mistral()
+
+        path_tekken_json = self.dir_model / "tekken.json"
+        path_tokenizer_json = self.dir_model / "tokenizer.json"
+        if path_tekken_json.is_file() and not path_tokenizer_json.is_file():
+            self._set_vocab_mistral()
+
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            try:
+                self._set_vocab_llama_hf()
+            except (FileNotFoundError, TypeError):
+                # Llama 3
+                self._set_vocab_gpt2()
+
+        # Apply to CodeLlama only (and ignore for Llama 3 with a vocab size of 128256)
+        if self.hparams.get("vocab_size", 32000) == 32016:
+            special_vocab = gguf.SpecialVocab(
+                self.dir_model, load_merges=False,
+                special_token_types = ['prefix', 'suffix', 'middle', 'eot']
+            )
+            special_vocab._set_special_token("prefix", 32007)
+            special_vocab._set_special_token("suffix", 32008)
+            special_vocab._set_special_token("middle", 32009)
+            special_vocab._set_special_token("eot",    32010)
+            special_vocab.add_to_gguf(self.gguf_writer)
+
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+        # Apply to granite small models only
+        if self.hparams.get("vocab_size", 32000) == 49152:
+            self.gguf_writer.add_add_bos_token(False)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        if not self.is_mistral_format:
+            self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
         if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
@@ -2033,13 +2097,25 @@ def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        n_head = self.hparams["num_attention_heads"]
-        n_kv_head = self.hparams.get("num_key_value_heads")
+        n_head = self.find_hparam(["n_heads", "num_attention_heads"])
+        n_kv_head = self.find_hparam(["n_kv_heads", "num_key_value_heads"])
+
+        vision_prefixes = [
+            "vision_encoder.",
+            "vision_language_adapter.",
+            "patch_merger.",
+            "pre_mm_projector_norm",
+        ]
+
         is_multimodal_tensor = "vision_tower" in name \
             or "vision_model" in name \
             or "audio_tower" in name \
             or "model.connector" in name \
-            or "multi_modal_projector" in name
+            or "multi_modal_projector" in name \
+            or any(
+                name.startswith(prefix)
+                for prefix in vision_prefixes
+            )
 
         if is_multimodal_tensor:
             return [] # skip vision tensors
@@ -2155,13 +2231,18 @@ class LlavaVisionModel(MmprojModel):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        if self.hparams["model_type"] == "pixtral":
+        if self.hparams.get("model_type") == "pixtral":
             # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py
             self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
             self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
-            logger.info(f"Image break token id: {self.img_break_tok_id}")
+        elif self.is_mistral_format:
+            # hparams is already vision config here so norm_eps is only defined in global_config.
+            self.hparams["norm_eps"] = self.global_config.get("norm_eps", None)
+            assert self.hparams["norm_eps"] is not None, "norm_eps not found in params.json"
+            self.img_break_tok_id = self.find_vparam(["image_break_token_id"])
         else:
             raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
+        logger.info(f"Image break token id: {self.img_break_tok_id}")
 
     def get_token_id(self, token: str) -> int:
         tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
@@ -2175,7 +2256,7 @@ def get_token_id(self, token: str) -> int:
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        if hparams["model_type"] == "pixtral":
+        if hparams.get("model_type") == "pixtral":
             self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PIXTRAL)
             self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
 
@@ -2193,18 +2274,30 @@ def set_gguf_parameters(self):
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
-        n_head = self.hparams["num_attention_heads"]
+        n_head = (
+            self.hparams["num_attention_heads"] if not self.is_mistral_format else self.find_vparam(["num_attention_heads"])
+        )
         n_kv_head = n_head
 
-        if name.startswith("multi_modal_projector.") or name.startswith("vision_tower."):
+        valid_prefixes = (
+            "multi_modal_projector.",
+            "vision_tower.",
+            "vision_encoder.",
+            "vision_language_adapter.",
+            "patch_merger.",
+            "pre_mm_projector_norm",
+        )
+
+        if any(name.startswith(prefix) for prefix in valid_prefixes):
             # process vision tensors
-            if name.endswith(("q_proj.weight", "q_proj.bias")):
+            if name.endswith(("q_proj.weight", "q_proj.bias")) and not self.is_mistral_format:
                 data_torch = LlamaModel.permute(data_torch, n_head, n_head)
-            if name.endswith(("k_proj.weight", "k_proj.bias")):
+            if name.endswith(("k_proj.weight", "k_proj.bias")) and not self.is_mistral_format:
                 data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
             return [(self.map_tensor_name(name), data_torch)]
 
-        if self.img_break_tok_id > 0 and "embed_tokens.weight" in name:
+        embed_key = "embed_tokens.weight" if not self.is_mistral_format else "tok_embeddings.weight"
+        if self.img_break_tok_id > 0 and embed_key in name:
             logger.info(f"Extracting [IMG_BREAK] token embedding from {name}")
             # for pixtral model, we need to extract the [IMG_BREAK] token embedding
             img_break_embd = data_torch[self.img_break_tok_id]
@@ -2233,10 +2326,9 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_vision_use_gelu(True)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, new_name, n_dims  # unused
         if ".embeddings." in name:
             return gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -3224,12 +3316,9 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6))
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, name, n_dims  # unused
-        if ".patch_embd." in new_name:
-            return gguf.GGMLQuantizationType.F16
         if ".position_embd." in new_name:
             return gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -3302,10 +3391,9 @@ def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         yield ("audio_tower.embed_positions.weight", pos_embd)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, new_name, n_dims  # unused
         if ".conv" in name and ".weight" in name:
             return gguf.GGMLQuantizationType.F16
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name.startswith("thinker."):
@@ -3351,12 +3439,9 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, name, n_dims  # unused
-        if ".patch_embd." in new_name:
-            return gguf.GGMLQuantizationType.F16
         if ".position_embd." in new_name:
             return gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def _mapping_interns1_name(self, name):
         names_map = {
@@ -3526,7 +3611,7 @@ class Qwen3MoeModel(Qwen2MoeModel):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        hparams = ModelBase.load_hparams(self.dir_model)
+        hparams = ModelBase.load_hparams(self.dir_model, False)
         self.origin_hf_arch = hparams.get('architectures', [None])[0]
 
     def set_vocab(self):
@@ -4683,7 +4768,7 @@ class NomicBertModel(BertModel):
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, **kwargs: Any):
         hparams = kwargs.pop("hparams", None)
         if hparams is None:
-            hparams = ModelBase.load_hparams(dir_model)
+            hparams = ModelBase.load_hparams(dir_model, False)
 
         self.is_moe = bool(hparams.get("moe_every_n_layers"))
         self.model_arch = gguf.MODEL_ARCH.NOMIC_BERT_MOE if self.is_moe else gguf.MODEL_ARCH.NOMIC_BERT
@@ -4990,13 +5075,12 @@ def set_gguf_parameters(self):
             self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, new_name, n_dims  # unused
         # related to https://github.com/ggml-org/llama.cpp/issues/13025
         if "input_projection" in name:
             return gguf.GGMLQuantizationType.F16
         if ".embeddings." in name:
             return gguf.GGMLQuantizationType.F32
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -5770,6 +5854,11 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("SeedOssForCausalLM")
+class SeedOssModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.SEED_OSS
+
+
 @ModelBase.register("Olmo2ForCausalLM")
 class Olmo2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.OLMO2
@@ -7655,10 +7744,9 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        del bid, new_name, n_dims  # unused
         if ".conv" in name and ".weight" in name:
             return gguf.GGMLQuantizationType.F16
-        return False
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -8179,8 +8267,7 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling.get("original_max_position_embeddings", 4096))
 
 
-@ModelBase.register("Lfm2ForCausalLM")
-@ModelBase.register("LFM2ForCausalLM")
+@ModelBase.register("Lfm2ForCausalLM", "LFM2ForCausalLM")
 class LFM2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.LFM2
 
@@ -8215,6 +8302,13 @@ def set_gguf_parameters(self):
         self._add_feed_forward_length()
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
+        if is_vision_tensor:
+            # skip vision tensors
+            return []
+
+        name = name.replace("language_model.", "")
+
         # conv op requires 2d tensor
         if 'conv.conv' in name:
             data_torch = data_torch.squeeze(1)
@@ -8222,6 +8316,41 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("Lfm2VlForConditionalGeneration")
+class LFM2VLModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        # TODO(tarek): for dynamic resolution image_size is not specified, setting here for compatibility
+        self.hparams_vision["image_size"] = 256
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.LFM2)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.find_vparam(["layer_norm_eps"]))
+        self.gguf_writer.add_vision_projector_scale_factor(self.global_config.get("downsample_factor", 2))
+        self.gguf_writer.add_vision_use_gelu(True)
+        # python notation, e.g. for vision_feature_layer == -1, we pick last layer -> vision_feature_layers_to_drop = 0
+        vision_feature_layers_to_drop = -(self.global_config.get("vision_feature_layer", -1) + 1)
+        self.gguf_writer.add_vision_block_count(self.find_vparam(self.n_block_keys) - vision_feature_layers_to_drop)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
+
+        if is_vision_tensor:
+            # remove "model." prefix
+            name = name.replace("model.vision_tower.", "vision_tower.")
+            name = name.replace("model.multi_modal_projector.", "multi_modal_projector.")
+
+            if "patch_embedding.weight" in name:
+                data_torch = data_torch.view(data_torch.shape[0], 16, 16, 3).permute(0, 3, 1, 2)
+
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return [] # skip other tensors
+
+
 @ModelBase.register("SmallThinkerForCausalLM")
 class SmallThinkerModel(TextModel):
     model_arch = gguf.MODEL_ARCH.SMALLTHINKER
@@ -8304,6 +8433,83 @@ def prepare_tensors(self):
             if len(experts) > 0:
                 raise ValueError(f"Unprocessed experts: {experts}")
 
+
+class MistralModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA
+    model_name = "Mistral"
+    hf_arch = ""
+    is_mistral_format = True
+    undo_permute = False
+
+    @staticmethod
+    def get_community_chat_template(vocab: MistralVocab, templates_dir: Path, is_mistral_format: bool):
+        assert TokenizerVersion is not None, "mistral_common is not installed"
+        assert isinstance(vocab.tokenizer, (Tekkenizer, SentencePieceTokenizer)), (
+            f"Expected Tekkenizer or SentencePieceTokenizer, got {type(vocab.tokenizer)}"
+        )
+
+        if vocab.tokenizer.version == TokenizerVersion.v1:
+            return "mistral-v1"
+        elif vocab.tokenizer.version == TokenizerVersion.v3 and vocab.tokenizer_type == MistralTokenizerType.spm:
+            return "mistral-v3"
+        elif vocab.tokenizer.version == TokenizerVersion.v3 and vocab.tokenizer_type == MistralTokenizerType.tekken:
+            return "mistral-v3-tekken"
+        elif vocab.tokenizer.version == TokenizerVersion.v7 and vocab.tokenizer_type == MistralTokenizerType.spm:
+            return "mistral-v7"
+        elif vocab.tokenizer.version == TokenizerVersion.v7 and vocab.tokenizer_type == MistralTokenizerType.tekken:
+            return "mistral-v7-tekken"
+        elif vocab.tokenizer.version == TokenizerVersion.v11:
+            template_file = "Mistral-Small-3.2-24B-Instruct-2506.jinja"
+        elif vocab.tokenizer.version == TokenizerVersion.v13:
+            template_file = "unsloth-mistral-Devstral-Small-2507.jinja"
+        else:
+            err_message = f"Unknown tokenizer type: {vocab.tokenizer_type} and version {vocab.tokenizer.version}"
+            if is_mistral_format:
+                err_message += (
+                    " . Please pass --disable-mistral-community-chat-template argument to the CLI "
+                    "if you want to skip this error and use the Mistral official `mistral-common` pre-processing library."
+                )
+            raise ValueError(err_message)
+
+        template_path = templates_dir / template_file
+        if not template_path.exists():
+            raise FileNotFoundError(f"Template file not found: {template_path}")
+
+        with open(template_path, "r", encoding="utf-8") as f:
+            template = f.read()
+
+        return template
+
+
+class PixtralModel(LlavaVisionModel):
+    model_name = "Pixtral"
+    hf_arch = ""
+    is_mistral_format = True
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PIXTRAL)
+
+        self.gguf_writer.add_vision_attention_layernorm_eps(
+            self.find_hparam(["norm_eps"])
+        )
+        self.gguf_writer.add_rope_freq_base(self.find_vparam(["rope_theta"]))
+
+        self.gguf_writer.add_vision_use_silu(True)
+
+        # spatial_merge_size
+        if self.find_vparam(["mm_projector_id"]) == "patch_merge":
+            self.gguf_writer.add_vision_spatial_merge_size(
+                self.find_vparam(["spatial_merge_size"])
+            )
+
+    def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str:
+        if name == "vision_language_adapter.w_in.weight":
+            return "mm.1.weight"
+        elif name == "vision_language_adapter.w_out.weight":
+            return "mm.2.weight"
+        return super().map_tensor_name(name, try_suffixes)
+
 ###### CONVERSION LOGIC ######
 
 
@@ -8454,6 +8660,17 @@ def parse_args() -> argparse.Namespace:
         "--mmproj", action="store_true",
         help="(Experimental) Export multimodal projector (mmproj) for vision models. This will only work on some vision models. A prefix 'mmproj-' will be added to the output file name.",
     )
+    parser.add_argument(
+        "--mistral-format", action="store_true",
+        help="Whether the model is stored following the Mistral format.",
+    )
+    parser.add_argument(
+        "--disable-mistral-community-chat-template", action="store_true",
+        help=(
+            "Whether to disable usage of Mistral community chat templates. If set, use the Mistral official `mistral-common` library for tokenization and detokenization of Mistral models. "
+            "Using `mistral-common` ensure correctness and zero-day support of tokenization for models converted from the Mistral format but requires to manually setup the tokenization server."
+        )
+    )
 
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
@@ -8559,17 +8776,26 @@ def main() -> None:
         if "mmproj" not in fname_out.name:
             fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-")
 
+    is_mistral_format = args.mistral_format
+    disable_mistral_community_chat_template = args.disable_mistral_community_chat_template
+
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
         model_type = ModelType.MMPROJ if args.mmproj else ModelType.TEXT
-        hparams = ModelBase.load_hparams(dir_model)
-        model_architecture = get_model_architecture(hparams, model_type)
-        logger.info(f"Model architecture: {model_architecture}")
-        try:
-            model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type)
-        except NotImplementedError:
-            logger.error(f"Model {model_architecture} is not supported")
-            sys.exit(1)
+        hparams = ModelBase.load_hparams(dir_model, is_mistral_format)
+        if not is_mistral_format:
+            model_architecture = get_model_architecture(hparams, model_type)
+            logger.info(f"Model architecture: {model_architecture}")
+            try:
+                model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type)
+            except NotImplementedError:
+                logger.error(f"Model {model_architecture} is not supported")
+                sys.exit(1)
+        elif args.mmproj:
+            assert hparams.get("vision_encoder") is not None, "This model does not support multimodal"
+            model_class = PixtralModel
+        else:
+            model_class = MistralModel
 
         model_instance = model_class(dir_model, output_type, fname_out,
                                      is_big_endian=args.bigendian, use_temp_file=args.use_temp_file,
@@ -8578,7 +8804,8 @@ def main() -> None:
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
-                                     remote_hf_model_id=hf_repo_id)
+                                     remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template
+                                     )
 
         if args.vocab_only:
             logger.info("Exporting model vocab...")
diff --git a/convert_lora_to_gguf.py b/convert_lora_to_gguf.py
index 00a6733cbd360..a67c0536a4128 100755
--- a/convert_lora_to_gguf.py
+++ b/convert_lora_to_gguf.py
@@ -340,7 +340,7 @@ def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
             sys.exit(1)
     else:
         logger.info(f"Loading base model: {dir_base_model.name}")
-        hparams = ModelBase.load_hparams(dir_base_model)
+        hparams = ModelBase.load_hparams(dir_base_model, False)
 
     with torch.inference_mode():
         try:
diff --git a/docs/build-s390x.md b/docs/build-s390x.md
index 4d5857753ae68..f3cdd63be3ece 100644
--- a/docs/build-s390x.md
+++ b/docs/build-s390x.md
@@ -76,6 +76,23 @@ cmake --build build --config Release -j $(nproc)
     cmake --build build --config Release -j $(nproc)
     ```
 
+## IBM zDNN Accelerator
+
+This provides acceleration using the IBM zAIU co-processor located in the Telum I and Telum II processors. Make sure to have the [IBM zDNN library](https://github.com/IBM/zDNN) installed.
+
+#### Compile from source from IBM
+
+You may find the official build instructions here: [Building and Installing zDNN](https://github.com/IBM/zDNN?tab=readme-ov-file#building-and-installing-zdnn)
+
+### Compilation
+
+```bash
+cmake -S . -B build             \
+    -DCMAKE_BUILD_TYPE=Release  \
+    -DGGML_ZDNN=ON
+cmake --build build --config Release -j$(nproc)
+```
+
 ## Getting GGUF Models
 
 All models need to be converted to Big-Endian. You can achieve this in three cases:
@@ -145,15 +162,15 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
 ### 1. SIMD Acceleration
 
-Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
+Only available in IBM z15/LinuxONE 3 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
 
 ### 2. NNPA Vector Intrinsics Acceleration
 
-Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned off by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
+Only available in IBM z16/LinuxONE 4 or later system with the `-DGGML_NNPA=ON` (turned off by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
 
-### 3. zDNN Accelerator
+### 3. zDNN Accelerator (WIP)
 
-_Only available in IBM z16 / LinuxONE 4 or later system. No support currently available._
+Only available in IBM z17/LinuxONE 5 or later system with the `-DGGML_ZDNN=ON` compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs will default back to CPU routines.
 
 ### 4. Spyre Accelerator
 
@@ -229,11 +246,12 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 
 ## Appendix A: Hardware Support Matrix
 
-|         | Support | Minimum Compiler Version |
-| ------- | ------- | ------------------------ |
-| IBM z15 | ✅      |                          |
-| IBM z16 | ✅      |                          |
-| IBM z17 | ✅      | GCC 15.1.0               |
+|          | Support | Minimum Compiler Version |
+| -------- | ------- | ------------------------ |
+| IBM z15  | ✅      |                          |
+| IBM z16  | ✅      |                          |
+| IBM z17  | ✅      | GCC 15.1.0               |
+| IBM zDNN | ✅      |                          |
 
 -   ✅ - supported and verified to run as intended
 -   🚫 - unsupported, we are unlikely able to provide support
@@ -242,13 +260,14 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 
 |            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
 | ---------- | ----------- | ---- | ---- | ----- |
-| FP32       | ✅          | ✅   | ❓   | ❓    |
+| FP32       | ✅          | ✅   | ✅   | ❓    |
 | FP16       | ✅          | ✅   | ❓   | ❓    |
 | BF16       | 🚫          | 🚫   | ❓   | ❓    |
 | Q4_0       | ✅          | ✅   | ❓   | ❓    |
 | Q4_1       | ✅          | ✅   | ❓   | ❓    |
-| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
-| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
+| MXFP4      | 🚫          | 🚫   | ❓   | ❓    |
+| Q5_0       | ✅          | ✅   | ❓   | ❓    |
+| Q5_1       | ✅          | ✅   | ❓   | ❓    |
 | Q8_0       | ✅          | ✅   | ❓   | ❓    |
 | Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
 | Q3_K       | ✅          | ✅   | ❓   | ❓    |
@@ -273,4 +292,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 25, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Aug 22, 2025.
diff --git a/docs/build.md b/docs/build.md
index dd486fe293546..b35a898ba9ba4 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -197,13 +197,12 @@ The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enab
 
 The following compilation options are also available to tweak performance:
 
-| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
-|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
-| GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
-| GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
-| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
-| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
+| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                                                                                                                      |
+|-------------------------------|------------------------|---------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
+| GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models. There may be issues with numerical overflows (except for CDNA and RDNA4) and memory use will be higher. Prompt processing may become faster on recent datacenter GPUs (the custom kernels were tuned primarily for RTX 3000/4000).                                            |
+| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                                                                                                                  |
+| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                                                                                                           |
 
 ## MUSA
 
diff --git a/docs/multimodal/MobileVLM.md b/docs/multimodal/MobileVLM.md
index 4f5eca6190657..3bfab9f3d2291 100644
--- a/docs/multimodal/MobileVLM.md
+++ b/docs/multimodal/MobileVLM.md
@@ -194,7 +194,7 @@ llama_print_timings:       total time =   44411.01 ms /   377 tokens
 ## Orin compile and run
 ### compile
 ```sh
-make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 GGML_CUDA_F16=1 -j 32
+make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 -j 32
 ```
 ### run on Orin
 ### case 1
diff --git a/docs/multimodal/minicpmo2.6.md b/docs/multimodal/minicpmo2.6.md
index d4edc97182cef..5e74058e5d54a 100644
--- a/docs/multimodal/minicpmo2.6.md
+++ b/docs/multimodal/minicpmo2.6.md
@@ -13,7 +13,7 @@ If there are differences in usage, please refer to the official build [documenta
 
 Clone llama.cpp:
 ```bash
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 ```
 
diff --git a/docs/multimodal/minicpmv2.6.md b/docs/multimodal/minicpmv2.6.md
index 1b0ff5969184f..bc874bbd8cd39 100644
--- a/docs/multimodal/minicpmv2.6.md
+++ b/docs/multimodal/minicpmv2.6.md
@@ -12,7 +12,7 @@ If there are differences in usage, please refer to the official build [documenta
 
 Clone llama.cpp:
 ```bash
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 ```
 
diff --git a/docs/ops.md b/docs/ops.md
index 1a474d904714d..9a81ca0a9770c 100644
--- a/docs/ops.md
+++ b/docs/ops.md
@@ -12,91 +12,92 @@ Legend:
 - 🟡 Partially supported by this backend
 - ❌ Not supported by this backend
 
-| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan |
-|-----------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ |
-|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 |
-|                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ |
-|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
-|                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 |
-|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ |
-|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ |
-|                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ |
-|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 |
-|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ |
-|                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 |
-|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ |
+| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan | zDNN |
+|-----------|------|------|------|------|------|------|------|------|------|
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
+|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ |
+|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
+|                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
+|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
+|                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
+|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
+|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
+|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
+|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
diff --git a/docs/ops/zDNN.csv b/docs/ops/zDNN.csv
new file mode 100644
index 0000000000000..bf633844f5ecf
--- /dev/null
+++ b/docs/ops/zDNN.csv
@@ -0,0 +1,8134 @@
+"backend_name","op_name","op_params","test_mode","supported","error_message","backend_reg_name"
+"zDNN","ABS","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","ABS","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","SGN","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","SGN","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","NEG","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","NEG","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","STEP","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","STEP","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","TANH","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","TANH","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","ELU","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","ELU","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","RELU","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","RELU","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","SIGMOID","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","SIGMOID","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","GELU","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","GELU","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","GELU_QUICK","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","GELU_QUICK","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","SILU","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","SILU","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","HARDSWISH","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","HARDSWISH","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","HARDSIGMOID","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","HARDSIGMOID","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","EXP","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","EXP","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","ABS","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","ABS","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","SGN","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","SGN","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","NEG","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","NEG","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","STEP","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","STEP","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","TANH","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","TANH","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","ELU","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","ELU","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","RELU","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","RELU","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","SIGMOID","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","SIGMOID","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","GELU","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","GELU","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","GELU_QUICK","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","GELU_QUICK","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","SILU","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","SILU","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","HARDSWISH","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","HARDSWISH","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","HARDSIGMOID","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","HARDSIGMOID","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","EXP","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","EXP","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","ABS","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","ABS","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","SGN","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","SGN","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","NEG","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","NEG","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","STEP","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","STEP","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","TANH","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","TANH","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","ELU","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","ELU","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","RELU","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","RELU","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","SIGMOID","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","SIGMOID","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","GELU","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","GELU","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","GELU_QUICK","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","GELU_QUICK","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","SILU","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","SILU","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","HARDSWISH","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","HARDSWISH","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","HARDSIGMOID","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","HARDSIGMOID","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","EXP","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","EXP","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","zDNN"
+"zDNN","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","zDNN"
+"zDNN","ABS","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","ABS","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","SGN","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","SGN","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","NEG","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","NEG","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","STEP","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","STEP","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","TANH","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","TANH","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","ELU","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","ELU","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","RELU","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","RELU","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","SIGMOID","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","SIGMOID","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","GELU","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","GELU","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","GELU_QUICK","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","GELU_QUICK","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","SILU","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","SILU","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","HARDSWISH","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","HARDSWISH","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","HARDSIGMOID","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","HARDSIGMOID","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","EXP","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","EXP","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","zDNN"
+"zDNN","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","REGLU","type=f16,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU","type=f16,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f16,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f16,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f16,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[128,2,2,2],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[5,7,11,13],v=0,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[128,2,2,2],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[5,7,11,13],v=0,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[128,2,2,2],v=0,split","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[5,7,11,13],v=0,split","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","REGLU","type=f32,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU","type=f32,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","SWIGLU","type=f32,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU_ERF","type=f32,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[128,2,2,2],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[5,7,11,13],v=1,swapped=0","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[128,2,2,2],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[5,7,11,13],v=1,swapped=1","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[128,2,2,2],v=1,split","support","0","no","zDNN"
+"zDNN","GEGLU_QUICK","type=f32,ne_a=[5,7,11,13],v=1,split","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f32,n=1,m=8,r=2,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f32,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f32,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f32,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f32,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f16,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f16,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f16,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=f16,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=bf16,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=bf16,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=bf16,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=bf16,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_0,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_0,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_0,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_0,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_1,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_1,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_1,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_1,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_0,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_0,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_0,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_0,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_1,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_1,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_1,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_1,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q8_0,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q8_0,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q8_0,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q8_0,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q2_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q2_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q2_K,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q2_K,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q3_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q3_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q3_K,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q3_K,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_K,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q4_K,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_K,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q5_K,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q6_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q6_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q6_K,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=q6_K,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_xxs,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_xxs,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_xxs,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_xxs,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_xs,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_xs,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_xs,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_xs,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_s,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_s,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_s,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq2_s,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq3_xxs,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq3_xxs,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq3_xxs,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq3_xxs,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq1_s,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq1_s,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq1_s,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq1_s,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq1_m,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq1_m,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq1_m,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq1_m,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq4_nl,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq4_nl,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq4_nl,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq4_nl,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq3_s,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq3_s,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq3_s,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq3_s,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq4_xs,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq4_xs,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq4_xs,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=iq4_xs,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=i32,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=i32,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=i32,n=256,m=5,r=4,b=7,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS","type=i32,n=256,m=5,r=4,b=7,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=f32,n=1,m=8,r=2,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=f32,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=f32,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=f16,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=f16,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=bf16,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=bf16,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q4_0,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q4_0,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q4_1,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q4_1,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q5_0,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q5_0,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q5_1,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q5_1,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q8_0,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q8_0,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q2_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q2_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q3_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q3_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q4_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q4_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q5_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q5_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q6_K,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=q6_K,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq2_xxs,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq2_xxs,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq2_xs,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq2_xs,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq2_s,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq2_s,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq3_xxs,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq3_xxs,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq1_s,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq1_s,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq1_m,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq1_m,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq4_nl,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq4_nl,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq3_s,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq3_s,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq4_xs,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=iq4_xs,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=i32,n=256,m=5,r=4,b=1,v=0","support","0","no","zDNN"
+"zDNN","GET_ROWS_BACK","type=i32,n=256,m=5,r=4,b=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[1,8,1,3],nr23=[1,1],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[3,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[31,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[33,5,1,1],nr23=[2,3],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[3,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[31,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[33,5,1,1],nr23=[2,3],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[3,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[31,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[33,5,1,7],nr23=[2,3],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[3,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[31,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f32,ne=[33,5,1,7],nr23=[2,3],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[3,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[31,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[33,5,1,1],nr23=[2,3],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[3,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[31,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[33,5,1,1],nr23=[2,3],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[3,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[31,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[33,5,1,7],nr23=[2,3],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[3,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[31,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=f16,ne=[33,5,1,7],nr23=[2,3],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[3,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[31,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[33,5,1,1],nr23=[2,3],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[3,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[31,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[33,5,1,1],nr23=[2,3],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[3,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[31,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[33,5,1,7],nr23=[2,3],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[3,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[31,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=bf16,ne=[33,5,1,7],nr23=[2,3],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[96,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[96,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[96,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_0,ne=[96,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[96,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[96,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[96,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_1,ne=[96,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[96,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[96,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[96,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_0,ne=[96,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[96,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[96,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[96,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_1,ne=[96,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[96,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[96,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[96,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q8_0,ne=[96,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q2_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q3_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q4_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q5_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=q6_K,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xxs,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_xs,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq2_s,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_xxs,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_s,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq1_m,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[96,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[96,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[96,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_nl,ne=[96,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq3_s,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[256,5,1,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[256,11,1,1],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[768,3,1,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[256,5,1,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[256,11,1,1],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[768,3,1,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[256,5,7,3],nr23=[1,1],r=1,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[256,11,1,7],nr23=[2,3],r=7,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[768,3,7,1],nr23=[2,3],r=2,v=0","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[256,5,7,3],nr23=[1,1],r=1,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[256,11,1,7],nr23=[2,3],r=7,v=1","support","0","no","zDNN"
+"zDNN","SET_ROWS","type=iq4_xs,ne=[768,3,7,1],nr23=[2,3],r=2,v=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=avg,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=1,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=1,s0=2,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=1,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=1,k1=3,s0=2,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=1,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=1,s0=2,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=1,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=1,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=1,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=1,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=1,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=0,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=0,p1=1","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=1,p1=0","support","0","no","zDNN"
+"zDNN","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=1,p1=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=1,s1=0,p0=0,p1=0,d0=1,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=1,s1=0,p0=0,p1=0,d0=3,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=1,s1=0,p0=3,p1=0,d0=1,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=1,s1=0,p0=3,p1=0,d0=3,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=3,s1=0,p0=0,p1=0,d0=1,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=3,s1=0,p0=0,p1=0,d0=3,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=3,s1=0,p0=3,p1=0,d0=1,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,2,2,1],ne_kernel=[3,2,2,1],s0=3,s1=0,p0=3,p1=0,d0=3,d1=0,is_2D=0","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[10,10,3,1],ne_kernel=[3,3,3,1],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=0,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=0,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=0,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=3,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=3,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=3,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=0,p1=3,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=0,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=0,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=0,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=3,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=3,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=3,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=1,p0=3,p1=3,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=0,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=0,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=0,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=3,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=3,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=3,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=0,p1=3,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=0,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=0,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=0,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=3,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=3,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=3,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=1,s1=3,p0=3,p1=3,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=0,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=0,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=0,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=3,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=3,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=3,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=0,p1=3,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=0,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=0,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=0,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=3,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=3,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=3,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=1,p0=3,p1=3,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=0,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=0,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=0,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=3,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=3,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=3,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=0,p1=3,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=0,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=0,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=0,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=3,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=3,d0=1,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=3,d0=3,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[20,20,2,2],ne_kernel=[3,3,2,2],s0=3,s1=3,p0=3,p1=3,d0=3,d1=3,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,32],ne_kernel=[3,3,1,32],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,32],ne_kernel=[3,3,2,32],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,1024],ne_kernel=[3,3,1,1024],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,1024],ne_kernel=[3,3,2,1024],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,2048],ne_kernel=[3,3,1,2048],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,2048],ne_kernel=[3,3,2,2048],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,2560],ne_kernel=[3,3,1,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,2560],ne_kernel=[3,3,2,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[5,5,1,32],ne_kernel=[3,4,1,32],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,1],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,1],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,12],type_kernel=f32,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,12],type_kernel=f16,stride0=1,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=2,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,1,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,2,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,1,2],ne_kernel=[11,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,3,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[1,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[2,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[3,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,1,2],ne_kernel=[11,11,1,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,1],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,1],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,1,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,2,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,1,25,2],ne_kernel=[11,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,3,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[1,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[2,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[1,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[3,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,12],type_kernel=f32,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D","ne_input=[141,133,25,2],ne_kernel=[11,11,25,12],type_kernel=f16,stride0=3,stride1=5,padding0=5,padding1=5,dilation0=2,dilation1=4,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D_DW","ne_input=[17,34,9,1],ne_kernel=[3,3,1,9],stride=1,padding=0,dilation=1,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D_DW","ne_input=[17,34,9,1],ne_kernel=[3,3,1,9],stride=1,padding=0,dilation=1,cwhn=1","support","0","no","zDNN"
+"zDNN","CONV_2D_DW","ne_input=[32,8,64,1],ne_kernel=[3,3,1,64],stride=2,padding=1,dilation=1,cwhn=0","support","0","no","zDNN"
+"zDNN","CONV_2D_DW","ne_input=[32,8,64,1],ne_kernel=[3,3,1,64],stride=2,padding=1,dilation=1,cwhn=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[3,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[3,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[3,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[3,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1337,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1337,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1337,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1337,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1337,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1337,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1337,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1337,1,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1337,1,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[3,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[3,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[3,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[3,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[3,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[3,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[3,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[3,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[3,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1337,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1337,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1337,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1337,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1337,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1337,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1337,1,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1337,1,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1337,1,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[3,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[3,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[3,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[3,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[3,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[3,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1337,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1337,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,1,1,1],ne_kernel=[1337,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1337,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1337,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[1337,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1337,9,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1337,9,1,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,1,1,1],ne_kernel=[1337,9,1,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[3,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[3,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[3,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[3,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[3,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[3,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[3,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[3,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[3,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1337,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1337,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[1,7,1,1],ne_kernel=[1337,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1337,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1337,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,7,1,1],ne_kernel=[1337,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1337,9,7,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1337,9,7,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[13,7,1,1],ne_kernel=[1337,9,7,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[197,32,1,1],ne_kernel=[16,32,32,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[3,2,1,1],ne_kernel=[2,3,2,1],s0=3,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[3,2,1,1],ne_kernel=[2,3,2,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[3,2,1,1],ne_kernel=[2,3,2,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[3,2,1,1],ne_kernel=[3,2,2,1],s0=2,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[3,2,1,1],ne_kernel=[3,2,2,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[3,2,1,1],ne_kernel=[3,1,2,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_2D","ne_input=[3,2,3,1],ne_kernel=[2,2,1,3],stride=1","support","0","no","zDNN"
+"zDNN","CONV_TRANSPOSE_2D","ne_input=[10,10,9,1],ne_kernel=[3,3,1,9],stride=2","support","0","no","zDNN"
+"zDNN","COUNT_EQUAL","type=f32,ne=[4,500,1,1]","support","0","no","zDNN"
+"zDNN","COUNT_EQUAL","type=f32,ne=[4,5000,1,1]","support","0","no","zDNN"
+"zDNN","ARGMAX","type=f32,ne=[32,1,1,1]","support","0","no","zDNN"
+"zDNN","ARGMAX","type=f32,ne=[100,10,1,1]","support","0","no","zDNN"
+"zDNN","ARGMAX","type=f32,ne=[1024,10,1,1]","support","0","no","zDNN"
+"zDNN","ARGMAX","type=f32,ne=[1024,12,1,1]","support","0","no","zDNN"
+"zDNN","ARGMAX","type=f32,ne=[2000,10,1,1]","support","0","no","zDNN"
+"zDNN","ARGMAX","type=f32,ne=[5438,3,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,1],nr=[2,1,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,2,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,2,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,2]","support","0","no","zDNN"
+"zDNN","REPEAT","type=i32,ne=[10,5,4,1],nr=[2,1,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=i16,ne=[10,5,4,1],nr=[1,1,1,2]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,3],nr=[2,1,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,2,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,2,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,2]","support","0","no","zDNN"
+"zDNN","REPEAT","type=i32,ne=[10,5,4,3],nr=[2,1,1,1]","support","0","no","zDNN"
+"zDNN","REPEAT","type=i16,ne=[10,5,4,3],nr=[1,1,1,2]","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,1,1,1],v=0","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[2,1,1,1],v=0","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,2,1,1],v=0","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,1,2,1],v=0","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,1,1,2],v=0","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,1,1,1],v=1","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[2,1,1,1],v=1","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,2,1,1],v=1","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,1,2,1],v=1","support","0","no","zDNN"
+"zDNN","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,1,1,2],v=1","support","0","no","zDNN"
+"zDNN","DUP","type=f32,ne=[10,10,20,1]","support","0","no","zDNN"
+"zDNN","DUP","type=f16,ne=[10,10,20,1]","support","0","no","zDNN"
+"zDNN","DUP","type=i32,ne=[10,10,20,1]","support","0","no","zDNN"
+"zDNN","DUP","type=i16,ne=[10,10,20,1]","support","0","no","zDNN"
+"zDNN","DUP","type=f32,ne=[10,10,5,1],permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","DUP","type=f16,ne=[10,10,5,1],permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","DUP","type=f32,ne=[10,10,5,1],permute=[1,0,2,3]","support","0","no","zDNN"
+"zDNN","DUP","type=f16,ne=[10,10,5,1],permute=[1,0,2,3]","support","0","no","zDNN"
+"zDNN","DUP","type=i16,ne=[10,8,3,1],permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","DUP","type=i16,ne=[10,8,3,1],permute=[1,2,0,3]","support","0","no","zDNN"
+"zDNN","SET","type_src=f32,type_dst=f32,ne=[6,5,4,3],dim=1","support","0","no","zDNN"
+"zDNN","SET","type_src=f32,type_dst=f32,ne=[6,5,4,3],dim=2","support","0","no","zDNN"
+"zDNN","SET","type_src=f32,type_dst=f32,ne=[6,5,4,3],dim=3","support","0","no","zDNN"
+"zDNN","SET","type_src=i32,type_dst=i32,ne=[6,5,4,3],dim=1","support","0","no","zDNN"
+"zDNN","SET","type_src=i32,type_dst=i32,ne=[6,5,4,3],dim=2","support","0","no","zDNN"
+"zDNN","SET","type_src=i32,type_dst=i32,ne=[6,5,4,3],dim=3","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[1,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[1,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[1,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[2,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[2,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[2,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[3,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[3,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[3,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[1,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[1,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[1,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[2,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[2,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[2,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[3,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[3,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[3,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[1,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[1,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[1,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[2,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[2,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[2,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[3,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[3,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[3,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[32,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[32,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[32,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[64,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[64,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[64,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[96,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[96,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=q4_0,ne=[96,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[32,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[32,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[32,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[64,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[64,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[64,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[96,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[96,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=q4_1,ne=[96,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[32,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[32,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[32,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[64,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[64,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[64,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[96,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[96,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=q5_0,ne=[96,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[32,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[32,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[32,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[64,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[64,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[64,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[96,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[96,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=q5_1,ne=[96,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[32,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[32,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[32,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[64,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[64,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[64,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[96,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[96,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=q8_0,ne=[96,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=q2_K,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=q3_K,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=q4_K,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=q5_K,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=q6_K,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=iq2_xxs,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=iq2_xs,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=iq2_s,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=iq3_xxs,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=iq1_s,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=iq1_m,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[32,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[32,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[32,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[64,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[64,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[64,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[96,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[96,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=iq4_nl,ne=[96,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=iq3_s,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[256,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[256,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[512,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[512,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[512,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[768,2,3,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[768,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=iq4_xs,ne=[768,2,3,4],permute_src=[0,3,1,2],permute_dst=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=bf16,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=bf16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q4_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q4_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q4_1,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q4_1,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q5_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q5_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q5_1,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q5_1,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q8_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q8_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q2_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q2_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q3_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q3_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q4_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q4_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q5_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q5_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q6_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=q6_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq2_xxs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq2_xxs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq2_xs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq2_xs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq2_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq2_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq3_xxs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq3_xxs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq1_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq1_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq1_m,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq1_m,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq4_nl,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq4_nl,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq3_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq3_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq4_xs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=iq4_xs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=f16,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=f16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=bf16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q4_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q4_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q4_1,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q4_1,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q5_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q5_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q5_1,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q5_1,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q8_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q8_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q2_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q2_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q3_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q3_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q4_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q4_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q5_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q5_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q6_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=q6_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq2_xxs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq2_xxs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq2_xs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq2_xs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq2_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq2_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq3_xxs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq3_xxs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq1_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq1_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq1_m,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq1_m,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq4_nl,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq4_nl,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq3_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq3_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq4_xs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=iq4_xs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f16,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=bf16,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=bf16,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q4_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q4_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q4_1,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q4_1,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q5_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q5_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q5_1,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q5_1,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q8_0,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q8_0,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q2_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q2_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q3_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q3_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q4_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q4_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q5_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q5_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q6_K,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=q6_K,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq2_xxs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq2_xxs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq2_xs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq2_xs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq2_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq2_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq3_xxs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq3_xxs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq1_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq1_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq1_m,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq1_m,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq4_nl,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq4_nl,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq3_s,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq3_s,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq4_xs,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=iq4_xs,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=bf16,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_0,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_1,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_0,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_1,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q8_0,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q2_K,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q3_K,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q4_K,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q5_K,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=q6_K,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xxs,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_xs,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq2_s,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_xxs,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_s,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq1_m,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_nl,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq3_s,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=f32,ne=[256,4,4,4],permute_src=[0,0,0,0],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=iq4_xs,type_dst=f32,ne=[256,2,3,4],permute_src=[0,2,1,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f16,ne=[256,2,3,4],permute_src=[1,0,2,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f16,type_dst=f32,ne=[256,2,3,4],permute_src=[1,0,2,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f16,ne=[256,2,3,4],permute_src=[1,0,2,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CPY","type_src=f32,type_dst=f32,ne=[256,2,3,4],permute_src=[1,0,2,3],permute_dst=[0,0,0,0]","support","0","no","zDNN"
+"zDNN","CONT","type=f32,ne=[10,10,10,1]","support","0","no","zDNN"
+"zDNN","CONT","type=f32,ne=[2,1,1,1]","support","0","no","zDNN"
+"zDNN","CONT","type=f32,ne=[2,1,3,5]","support","0","no","zDNN"
+"zDNN","CONT","type=f32,ne=[2,3,5,7]","support","0","no","zDNN"
+"zDNN","CONT","type=f16,ne=[2,1,1,1]","support","0","no","zDNN"
+"zDNN","CONT","type=f16,ne=[2,1,3,5]","support","0","no","zDNN"
+"zDNN","CONT","type=f16,ne=[2,3,5,7]","support","0","no","zDNN"
+"zDNN","CONT","type=bf16,ne=[2,1,1,1]","support","0","no","zDNN"
+"zDNN","CONT","type=bf16,ne=[2,1,3,5]","support","0","no","zDNN"
+"zDNN","CONT","type=bf16,ne=[2,3,5,7]","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,1,1],nr=[32,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,1,1],nr=[32,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,1,1],nr=[32,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,1,1],nr=[32,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,320,320],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,320,320],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,320,320],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,320,320],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,3],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,3],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,3],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,3],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,3],nr=[2,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,3],nr=[2,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,3],nr=[2,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,3],nr=[2,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,3],nr=[1,2,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,3],nr=[1,2,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,3],nr=[1,2,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,3],nr=[1,2,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,3],nr=[1,1,2,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,3],nr=[1,1,2,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,3],nr=[1,1,2,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,3],nr=[1,1,2,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,3],nr=[1,1,1,2],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,3],nr=[1,1,1,2],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,3],nr=[1,1,1,2],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,3],nr=[1,1,1,2],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,3],nr=[1,1,2,2],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,3],nr=[1,1,2,2],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,3],nr=[1,1,2,2],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,3],nr=[1,1,2,2],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,3],nr=[1,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,3],nr=[1,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,3],nr=[1,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,3],nr=[1,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[10,5,4,3],nr=[2,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[10,5,4,3],nr=[2,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[10,5,4,3],nr=[2,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[10,5,4,3],nr=[2,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1280,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1280,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1280,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1280,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1280,1,1,1],nr=[1,16,16,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1280,1,1,1],nr=[1,16,16,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1280,1,1,1],nr=[1,16,16,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1280,1,1,1],nr=[1,16,16,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1280,16,16,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1280,16,16,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1280,16,16,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1280,16,16,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1280,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1280,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1280,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1280,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,1280,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,1280,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,1280,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,1280,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[16,16,1280,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[16,16,1280,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[16,16,1280,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[16,16,1280,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,1920,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,1920,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,1920,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,1920,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,2560,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,2560,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,2560,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,2560,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,1280,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,1280,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,1280,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,1280,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,1920,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,1920,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,1920,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,1920,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[1,1,640,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[1,1,640,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[1,1,640,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[1,1,640,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[5120,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[5120,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[5120,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[5120,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f16,ne=[640,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f16,ne=[640,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f16,ne=[640,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f16,ne=[640,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,1,1],nr=[32,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,1,1],nr=[32,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,1,1],nr=[32,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,1,1],nr=[32,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,320,320],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,320,320],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,320,320],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,320,320],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,3],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,3],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,3],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[2,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,3],nr=[2,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,3],nr=[2,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,3],nr=[2,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,2,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,3],nr=[1,2,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,3],nr=[1,2,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,3],nr=[1,2,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,1,2,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,3],nr=[1,1,2,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,3],nr=[1,1,2,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,3],nr=[1,1,2,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,1,1,2],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,3],nr=[1,1,1,2],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,3],nr=[1,1,1,2],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,3],nr=[1,1,1,2],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,1,2,2],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,3],nr=[1,1,2,2],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,3],nr=[1,1,2,2],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,3],nr=[1,1,2,2],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,3],nr=[1,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,3],nr=[1,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,3],nr=[1,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[2,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[10,5,4,3],nr=[2,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[10,5,4,3],nr=[2,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[10,5,4,3],nr=[2,2,2,2],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1280,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1280,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1280,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1280,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1280,1,1,1],nr=[1,16,16,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1280,1,1,1],nr=[1,16,16,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1280,1,1,1],nr=[1,16,16,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1280,1,1,1],nr=[1,16,16,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1280,16,16,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1280,16,16,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1280,16,16,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1280,16,16,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1280,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1280,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1280,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1280,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,1280,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,1280,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,1280,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,1280,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[16,16,1280,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[16,16,1280,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[16,16,1280,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[16,16,1280,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,1920,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,1920,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,1920,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,1920,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,2560,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,2560,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,2560,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,2560,1],nr=[16,16,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,1280,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,1280,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,1280,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,1280,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,1920,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,1920,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,1920,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,1920,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[1,1,640,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[1,1,640,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[1,1,640,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[1,1,640,1],nr=[32,32,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[5120,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[5120,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[5120,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[5120,1,1,1],nr=[1,256,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[640,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","SUB","type=f32,ne=[640,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","MUL","type=f32,ne=[640,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","DIV","type=f32,ne=[640,1,1,1],nr=[1,1,1,1],nf=1","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[2,1,1,1],nf=2","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[16,5,4,3],nr=[1,2,1,1],nf=3","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,1,2,1],nf=4","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[16,5,4,3],nr=[1,1,1,2],nf=5","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,1,2,2],nf=6","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[10,5,4,3],nr=[1,2,2,2],nf=7","support","0","no","zDNN"
+"zDNN","ADD","type=f32,ne=[16,5,4,3],nr=[2,2,2,2],nf=8","support","0","no","zDNN"
+"zDNN","ADD1","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=0.000000","support","0","no","zDNN"
+"zDNN","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=1.000000","support","0","no","zDNN"
+"zDNN","SOFTCAP","type=f32,ne=[10,10,10,10],softcap=50.000000","support","0","no","zDNN"
+"zDNN","SILU_BACK","type=f32,ne=[64,5,4,3],eps=0.000001","support","0","no","zDNN"
+"zDNN","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000000","support","0","no","zDNN"
+"zDNN","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000000","support","0","no","zDNN"
+"zDNN","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000000","support","0","no","zDNN"
+"zDNN","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000000","support","0","no","zDNN"
+"zDNN","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000000","support","0","no","zDNN"
+"zDNN","L2_NORM","type=f32,ne=[64,5,4,3]","support","0","no","zDNN"
+"zDNN","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001","support","0","no","zDNN"
+"zDNN","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001","support","0","no","zDNN"
+"zDNN","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001","support","0","no","zDNN"
+"zDNN","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001","support","0","no","zDNN"
+"zDNN","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000001","support","0","no","zDNN"
+"zDNN","L2_NORM","type=f32,ne=[64,5,4,3]","support","0","no","zDNN"
+"zDNN","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100","support","0","no","zDNN"
+"zDNN","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100","support","0","no","zDNN"
+"zDNN","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100","support","0","no","zDNN"
+"zDNN","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100","support","0","no","zDNN"
+"zDNN","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000100","support","0","no","zDNN"
+"zDNN","L2_NORM","type=f32,ne=[64,5,4,3]","support","0","no","zDNN"
+"zDNN","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000","support","0","no","zDNN"
+"zDNN","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000","support","0","no","zDNN"
+"zDNN","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000","support","0","no","zDNN"
+"zDNN","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000","support","0","no","zDNN"
+"zDNN","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.100000","support","0","no","zDNN"
+"zDNN","L2_NORM","type=f32,ne=[64,5,4,3]","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.000000,broadcast=0","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.000000,broadcast=1","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.000001,broadcast=0","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.000001,broadcast=1","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.000100,broadcast=0","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.000100,broadcast=1","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.100000,broadcast=0","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.100000,broadcast=1","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=1.000000,broadcast=0","support","0","no","zDNN"
+"zDNN","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=1.000000,broadcast=1","support","0","no","zDNN"
+"zDNN","L2_NORM","type=f32,ne=[64,5,4,3]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[4,1536,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[4,1536,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[4,1536,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","zDNN"
+"zDNN","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","0","no","zDNN"
+"zDNN","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","0","no","zDNN"
+"zDNN","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","zDNN"
+"zDNN","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","zDNN"
+"zDNN","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=1,n_seqs=1","support","0","no","zDNN"
+"zDNN","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=1","support","0","no","zDNN"
+"zDNN","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=4","support","0","no","zDNN"
+"zDNN","RWKV_WKV6","type=f32,head_count=32,head_size=64,n_seq_tokens=128,n_seqs=4","support","0","no","zDNN"
+"zDNN","RWKV_WKV7","type=f32,head_count=32,head_size=64,n_seq_tokens=1,n_seqs=1","support","0","no","zDNN"
+"zDNN","RWKV_WKV7","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=1","support","0","no","zDNN"
+"zDNN","RWKV_WKV7","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=4","support","0","no","zDNN"
+"zDNN","RWKV_WKV7","type=f32,head_count=32,head_size=64,n_seq_tokens=128,n_seqs=4","support","0","no","zDNN"
+"zDNN","GATED_LINEAR_ATTN","type=f32,head_count=32,head_size=64,n_seq_tokens=1,n_seqs=1","support","0","no","zDNN"
+"zDNN","GATED_LINEAR_ATTN","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=1","support","0","no","zDNN"
+"zDNN","GATED_LINEAR_ATTN","type=f32,head_count=32,head_size=64,n_seq_tokens=32,n_seqs=4","support","0","no","zDNN"
+"zDNN","GATED_LINEAR_ATTN","type=f32,head_count=32,head_size=64,n_seq_tokens=128,n_seqs=4","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=4,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=5,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=6,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=7,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=8,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=9,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","1","yes","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f16,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=8,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=4,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f16,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f16,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f32,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_0,type_b=f16,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f16,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f32,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_K,type_b=f16,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[1,1],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[3,1],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[1,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[3,2],nr=[2,2],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=8,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,1,3,2],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=256,bs=[2,3],nr=[1,1],per=[0,3,2,1],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=1,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=8,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xxs,type_b=f16,m=16,n=16,k=1024,bs=[3,2],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=32,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q4_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=1,k=32,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=1,k=32,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=32,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q8_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q2_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q3_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q5_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=q6_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_xs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq2_s,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_s,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq1_m,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=1,k=32,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_nl,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq3_s,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=iq4_xs,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=1,k=1,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=64,n=2,k=128,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=83,n=2,k=128,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=64,n=2,k=64,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=83,n=2,k=64,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=64,n=45,k=128,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=45,k=64,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=193,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=67,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1057,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=129,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1057,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=129,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1057,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1057,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=129,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1057,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=129,bs=[1,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1057,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1057,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=129,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1057,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=129,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1057,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=128,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1056,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=128,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1056,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=128,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1056,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1057,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=129,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1057,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=129,bs=[1,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1057,bs=[1,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1056,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=128,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1056,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1057,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=129,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1057,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=129,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1057,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=128,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1056,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=128,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1056,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=128,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1056,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1057,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=129,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1057,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=129,bs=[2,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1057,bs=[2,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1056,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=128,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1056,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1057,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=129,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1057,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=129,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1057,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=128,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1056,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=128,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1056,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=128,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1056,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1057,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=129,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1057,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=129,bs=[2,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1057,bs=[2,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1056,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=128,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1056,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1057,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=129,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1057,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=129,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1057,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=128,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1056,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=128,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1056,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=128,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1056,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1057,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=129,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1057,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=129,bs=[4,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1057,bs=[4,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1056,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=128,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1056,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1057,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=129,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1057,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=129,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1057,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=128,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1056,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=128,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1056,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=128,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1056,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1057,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=129,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1057,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=129,bs=[4,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1057,bs=[4,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1056,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=128,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1056,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1057,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=129,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1057,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=129,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1057,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=128,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1056,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=128,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1056,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=128,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1056,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1057,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=129,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1057,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=129,bs=[8,1],nr=[1,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1057,bs=[8,1],nr=[1,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1056,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=128,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1056,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1057,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=129,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=128,n=1,k=1057,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1056,n=1,k=129,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=128,n=1,k=1057,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=128,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1056,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=128,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1056,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=128,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1056,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f16,type_b=f32,m=129,n=1,k=1057,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=1057,n=1,k=129,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=bf16,type_b=f32,m=129,n=1,k=1057,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=1057,n=1,k=129,bs=[8,1],nr=[4,1],per=[0,2,1,3],v=0","support","0","no","zDNN"
+"zDNN","MUL_MAT","type_a=f32,type_b=f32,m=129,n=1,k=1057,bs=[8,1],nr=[4,1],per=[0,1,2,3],v=1","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=16,n_used=16,b=0,m=32,n=1024,k=16","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=16,n_used=16,b=1,m=32,n=1024,k=16","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f32,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=f16,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_0,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_K,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=4,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=1,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=1,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=2,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=2,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=4,b=0,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xxs,type_b=f32,n_mats=8,n_used=4,b=1,m=512,n=129,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q4_1,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q5_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q5_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q5_1,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q5_1,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q8_0,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q2_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q2_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q3_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q3_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q5_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q5_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q6_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=q6_K,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_xs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_s,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq2_s,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq3_xxs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq3_xxs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq1_s,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq1_s,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq1_m,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq1_m,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq4_nl,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq4_nl,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq3_s,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq3_s,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq4_xs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=iq4_xs,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=bf16,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=1,k=256","support","0","no","zDNN"
+"zDNN","MUL_MAT_ID","type_a=bf16,type_b=f32,n_mats=4,n_used=2,b=0,m=512,n=32,k=256","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f32,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=f16,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q8_0,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_0,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_1,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=q4_K,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f32,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=1,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=1,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[1,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[1,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[3,1],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[1,2],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,1],trans_b=0","support","0","no","zDNN"
+"zDNN","OUT_PROD","type_a=iq2_xxs,type_b=f16,m=256,n=16,k=16,bs=[3,3],nr=[2,2],trans_b=0","support","0","no","zDNN"
+"zDNN","SQR","type=f16,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","SQRT","type=f16,ne=[10,3,3,2]","support","0","no","zDNN"
+"zDNN","LOG","type=f16,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","SIN","type=f16,ne=[10,2,2,2]","support","0","no","zDNN"
+"zDNN","COS","type=f16,ne=[10,2,2,2]","support","0","no","zDNN"
+"zDNN","CLAMP","type=f16,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","0","no","zDNN"
+"zDNN","SQR","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","SQRT","type=f32,ne=[10,3,3,2]","support","0","no","zDNN"
+"zDNN","LOG","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","SIN","type=f32,ne=[10,2,2,2]","support","0","no","zDNN"
+"zDNN","COS","type=f32,ne=[10,2,2,2]","support","0","no","zDNN"
+"zDNN","CLAMP","type=f32,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","0","no","zDNN"
+"zDNN","DIAG_MASK_INF","type=f32,ne=[10,10,1,1],n_past=5","support","0","no","zDNN"
+"zDNN","DIAG_MASK_INF","type=f32,ne=[10,10,3,1],n_past=5","support","0","no","zDNN"
+"zDNN","DIAG_MASK_INF","type=f32,ne=[10,10,3,2],n_past=5","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,3],mask=1,m_prec=f32,nr23=[3,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f32,nr23=[2,3],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,3],mask=1,m_prec=f16,nr23=[3,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f16,nr23=[2,3],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,3],mask=1,m_prec=f32,nr23=[3,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f32,nr23=[2,3],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,3],mask=1,m_prec=f16,nr23=[3,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f16,nr23=[2,3],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,3],mask=1,m_prec=f32,nr23=[3,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f32,nr23=[2,3],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,3],mask=1,m_prec=f16,nr23=[3,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f16,nr23=[2,3],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,3],mask=1,m_prec=f32,nr23=[3,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f32,nr23=[2,3],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,16,1,3],mask=1,m_prec=f16,nr23=[3,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,15,1,1],mask=1,m_prec=f16,nr23=[2,3],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,1024,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[15,1023,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,16,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,15,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1024,1024,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[1023,1023,1,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,2,32,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,2,32,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[16,2,32,1],mask=0,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[16,16,1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[15,15,1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[16,1024,1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[15,1023,1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1024,16,1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1023,15,1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1024,1024,1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1023,1023,1,1],scale=1.000000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[16,16,1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[15,15,1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[16,1024,1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[15,1023,1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1024,16,1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1023,15,1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1024,1024,1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1023,1023,1,1],scale=0.100000,max_bias=0.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[16,16,1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[15,15,1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[16,1024,1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[15,1023,1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1024,16,1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1023,15,1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1024,1024,1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1023,1023,1,1],scale=1.000000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[16,16,1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[15,15,1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[16,1024,1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[15,1023,1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1024,16,1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1023,15,1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1024,1024,1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","SOFT_MAX_BACK","type=f32,ne=[1023,1023,1,1],scale=0.100000,max_bias=8.000000","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.000000,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.000000,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","no","zDNN"
+"zDNN","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","no","zDNN"
+"zDNN","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","no","zDNN"
+"zDNN","ARGSORT","type=f32,ne=[8,1,1,1],order=0","support","0","no","zDNN"
+"zDNN","ARGSORT","type=f32,ne=[16,10,10,10],order=0","support","0","no","zDNN"
+"zDNN","ARGSORT","type=f32,ne=[60,10,10,10],order=0","support","0","no","zDNN"
+"zDNN","ARGSORT","type=f32,ne=[8,1,1,1],order=1","support","0","no","zDNN"
+"zDNN","ARGSORT","type=f32,ne=[16,10,10,10],order=1","support","0","no","zDNN"
+"zDNN","ARGSORT","type=f32,ne=[60,10,10,10],order=1","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=0","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=0","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=0","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=1","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=1","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=1","support","0","no","zDNN"
+"zDNN","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=257","support","0","no","zDNN"
+"zDNN","SUM","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","SUM_ROWS","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","MEAN","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","GROUP_NORM","type=f32,ne=[64,64,320,1],num_groups=32,eps=0.000001","support","0","no","zDNN"
+"zDNN","GROUP_NORM","type=f32,ne=[9,9,1280,1],num_groups=32,eps=0.000001","support","0","no","zDNN"
+"zDNN","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1]","support","0","no","zDNN"
+"zDNN","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1","support","0","no","zDNN"
+"zDNN","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","0","no","zDNN"
+"zDNN","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","0","no","zDNN"
+"zDNN","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","0","no","zDNN"
+"zDNN","TIMESTEP_EMBEDDING","type=f32,ne_a=[2,1,1,1],dim=320,max_period=10000","support","0","no","zDNN"
+"zDNN","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[1,3],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=64,hsv=64,nh=4,nr23=[4,3],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=80,hsv=80,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=192,hsv=192,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[1,1],kv=1024,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=1,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=3,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=32,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","FLASH_ATTN_EXT","hsk=576,hsv=512,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","zDNN"
+"zDNN","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","0","no","zDNN"
+"zDNN","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
+"zDNN","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","zDNN"
+"zDNN","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","0","no","zDNN"
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 11ff38762b848..abc4fa1c8931f 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -34,6 +34,7 @@ else()
     add_subdirectory(gen-docs)
     add_subdirectory(training)
     add_subdirectory(diffusion)
+    add_subdirectory(model-conversion)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading
diff --git a/examples/batched.swift/README.md b/examples/batched.swift/README.md
index 7f2e2fcdcf4a7..f08901506f94f 100644
--- a/examples/batched.swift/README.md
+++ b/examples/batched.swift/README.md
@@ -1,4 +1,5 @@
 This is a swift clone of `examples/batched`.
 
-$ `make`
-$ `./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]`
+```bash
+$ ./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]
+```
diff --git a/examples/eval-callback/eval-callback.cpp b/examples/eval-callback/eval-callback.cpp
index 4afd80eb454ad..61eefc7248a7e 100644
--- a/examples/eval-callback/eval-callback.cpp
+++ b/examples/eval-callback/eval-callback.cpp
@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <numeric>
 
 /**
  * This the arbitrary data which will be passed to each callback.
@@ -77,6 +78,12 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
         LOG("                                     ]\n");
         LOG("                                     sum = %f\n", sum);
     }
+
+    // TODO: make this abort configurable/optional?
+    if (std::isnan(sum)) {
+        LOG_ERR("encountered NaN - aborting\n");
+        exit(0);
+    }
 }
 
 /**
diff --git a/examples/llama.vim b/examples/llama.vim
index af3fd3935d765..736802d365541 100644
--- a/examples/llama.vim
+++ b/examples/llama.vim
@@ -17,7 +17,7 @@
 "
 " start the llama.cpp server with a FIM-compatible model. for example:
 "
-"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
+"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa --ubatch-size 512 --batch-size 1024 --cache-reuse 256
 "
 "   --batch-size [512, model max context]
 "
diff --git a/examples/lookahead/README.md b/examples/lookahead/README.md
index aab3cd0ca49b9..c82de2a5a9715 100644
--- a/examples/lookahead/README.md
+++ b/examples/lookahead/README.md
@@ -5,3 +5,9 @@ Demonstration of lookahead decoding technique:
 https://lmsys.org/blog/2023-11-21-lookahead-decoding/
 
 More info: https://github.com/ggml-org/llama.cpp/pull/4207
+
+Sample command:
+
+```bash
+llama-lookahead -hf ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF -p "// network server implemented in C\n// author: Peter Hacker\n\n#include" -e -ngl 99 -t 4 -n 512 -c 4096 -kvu
+```
diff --git a/examples/model-conversion/.gitignore b/examples/model-conversion/.gitignore
new file mode 100644
index 0000000000000..451227547fcc1
--- /dev/null
+++ b/examples/model-conversion/.gitignore
@@ -0,0 +1,3 @@
+.model_name
+data
+ppl
diff --git a/examples/model-conversion/CMakeLists.txt b/examples/model-conversion/CMakeLists.txt
new file mode 100644
index 0000000000000..fc1746ce4500c
--- /dev/null
+++ b/examples/model-conversion/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET llama-logits)
+add_executable(${TARGET} logits.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
diff --git a/examples/model-conversion/Makefile b/examples/model-conversion/Makefile
new file mode 100644
index 0000000000000..27d95b4f2bf5e
--- /dev/null
+++ b/examples/model-conversion/Makefile
@@ -0,0 +1,163 @@
+# Validation functions
+define validate_model_path
+	@if [ -z "$(MODEL_PATH)" ]; then \
+		echo "Error: MODEL_PATH must be provided either as:"; \
+		echo "  1. Environment variable: export MODEL_PATH=/path/to/model"; \
+		echo "  2. Command line argument: make $(1) MODEL_PATH=/path/to/model"; \
+		exit 1; \
+	fi
+endef
+
+define validate_embedding_model_path
+	@if [ -z "$(EMBEDDING_MODEL_PATH)" ]; then \
+		echo "Error: EMBEDDING_MODEL_PATH must be provided either as:"; \
+		echo "  1. Environment variable: export EMBEDDING_MODEL_PATH=/path/to/model"; \
+		echo "  2. Command line argument: make $(1) EMBEDDING_MODEL_PATH=/path/to/model"; \
+		exit 1; \
+	fi
+endef
+
+###
+### Casual Model targets/recipes
+###
+causal-convert-model-bf16: OUTTYPE=bf16
+causal-convert-model-bf16: causal-convert-model
+
+causal-convert-model:
+	$(call validate_model_path,causal-convert-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/causal/convert-model.sh
+
+causal-run-original-model:
+	$(call validate_model_path,causal-run-original-model)
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/run-org-model.py
+
+causal-run-converted-model:
+	@CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/causal/run-converted-model.sh
+
+causal-verify-logits: causal-run-original-model causal-run-converted-model
+	@./scripts/causal/compare-logits.py
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/utils/check-nmse.py -m ${MODEL_PATH}
+
+causal-run-original-embeddings:
+	@./scripts/causal/run-casual-gen-embeddings-org.sh
+
+causal-run-converted-embeddings:
+	@./scripts/causal/run-converted-model-embeddings-logits.sh
+
+causal-verify-embeddings: causal-run-original-embeddings causal-run-converted-embeddings
+	@./scripts/causal/compare-embeddings-logits.sh
+
+causal-inspect-original-model:
+	@./scripts/utils/inspect-org-model.py
+
+causal-inspect-converted-model:
+	@./scripts/utils/inspect-converted-model.sh
+
+causal-start-embedding-server:
+	@./scripts/utils/run-embedding-server.sh ${CONVERTED_MODEL}
+
+causal-curl-embedding-endpoint: causal-run-original-embeddings
+	@./scripts/utils/curl-embedding-server.sh | ./scripts/causal/compare-embeddings-logits.sh
+
+causal-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
+causal-quantize-Q8_0: causal-quantize-model
+
+causal-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
+causal-quantize-Q4_0: causal-quantize-model
+
+causal-quantize-model:
+	@CONVERTED_MODEL="$(CONVERTED_MODEL)" QUANTIZED_TYPE="$(QUANTIZED_TYPE)" ./scripts/utils/quantize.sh ${CONVERTED_MODEL} ${QUANTIZED_TYPE}
+	@echo "Export the quantized model path to QUANTIZED_MODEL variable in your environment"
+
+causal-run-quantized-model:
+	@QUANTIZED_MODEL="$(QUANTIZED_MODEL)" ./scripts/causal/run-converted-model.sh ${QUANTIZED_MODEL}
+
+
+###
+### Embedding Model targets/recipes
+###
+
+embedding-convert-model-bf16: OUTTYPE=bf16
+embedding-convert-model-bf16: embedding-convert-model
+
+embedding-convert-model:
+	$(call validate_embedding_model_path,embedding-convert-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/embedding/convert-model.sh
+
+embedding-run-original-model:
+	$(call validate_embedding_model_path,embedding-run-original-model)
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/embedding/run-original-model.py
+
+embedding-run-converted-model:
+	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/embedding/run-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
+	@./scripts/embedding/compare-embeddings-logits.sh
+
+embedding-inspect-original-model:
+	$(call validate_embedding_model_path,embedding-inspect-original-model)
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH}
+
+embedding-inspect-converted-model:
+	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/utils/inspect-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-start-embedding-server:
+	@./scripts/utils/run-embedding-server.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-curl-embedding-endpoint:
+	@./scripts/utils/curl-embedding-server.sh | ./scripts/embedding/compare-embeddings-logits.sh
+
+embedding-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
+embedding-quantize-Q8_0: embedding-quantize-model
+
+embedding-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
+embedding-quantize-Q4_0: embedding-quantize-model
+
+embedding-quantize-model:
+	@./scripts/utils/quantize.sh ${CONVERTED_EMBEDDING_MODEL} ${QUANTIZED_TYPE}
+	@echo "Export the quantized model path to QUANTIZED_EMBEDDING_MODEL variable in your environment"
+
+embedding-run-quantized-model:
+	@./scripts/embedding/run-converted-model.sh ${QUANTIZED_EMBEDDING_MODEL}
+
+###
+### Perplexity targets/recipes
+###
+perplexity-data-gen:
+	CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/utils/perplexity-gen.sh
+
+perplexity-run-full:
+	QUANTIZED_MODEL="$(QUANTIZED_MODEL)" LOOGITS_FILE="$(LOGITS_FILE)" \
+	./scripts/utils/perplexity-run.sh
+
+perplexity-run:
+	QUANTIZED_MODEL="$(QUANTIZED_MODEL)" ./scripts/utils/perplexity-run-simple.sh
+
+###
+### HuggingFace targets/recipes
+###
+
+hf-create-model:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}"
+
+hf-create-model-private:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -p
+
+hf-upload-gguf-to-model:
+	@./scripts/utils/hf-upload-gguf-model.py -m "${MODEL_PATH}" -r "${REPO_ID}" -o "${NAME_IN_REPO}"
+
+hf-create-collection:
+	@./scripts/utils/hf-create-collection.py -n "${NAME}" -d "${DESCRIPTION}" -ns "${NAMESPACE}"
+
+hf-add-model-to-collection:
+	@./scripts/utils/hf-add-model-to-collection.py -c "${COLLECTION}" -m "${MODEL}"
+
+
+.PHONY: clean
+clean:
+	@${RM} -rf data .converted_embedding_model.txt .converted_model.txt .embedding_model_name.txt .model_name.txt
+
diff --git a/examples/model-conversion/README.md b/examples/model-conversion/README.md
new file mode 100644
index 0000000000000..c924a6be3cd26
--- /dev/null
+++ b/examples/model-conversion/README.md
@@ -0,0 +1,335 @@
+# Model Conversion Example
+This directory contains scripts and code to help in the process of converting
+HuggingFace PyTorch models to GGUF format.
+
+The motivation for having this is that the conversion process can often be an
+iterative process, where the original model is inspected, converted, updates
+made to llama.cpp, converted again, etc. Once the model has been converted it
+needs to be verified against the original model, and then optionally quantified,
+and in some cases perplexity checked of the quantized model. And finally the
+model/models need to the ggml-org on Hugging Face. This tool/example tries to
+help with this process.
+
+### Overview
+The idea is that the makefile targets and scripts here can be used in the
+development/conversion process assisting with things like:
+
+* inspect/run the original model to figure out how it works
+* convert the original model to GGUF format
+* inspect/run the converted model
+* verify the logits produced by the original model and the converted model
+* quantize the model to GGUF format
+* run perplexity evaluation to verify that the quantized model is performing
+  as expected
+* upload the model to HuggingFace to make it available for others
+
+## Setup
+Create virtual python environment
+```console
+$ python3.11 -m venv venv
+$ source venv/bin/activate
+(venv) $ pip install -r requirements.txt
+```
+
+## Causal Language Model Conversion
+This section describes the steps to convert a causal language model to GGUF and
+to verify that the conversion was successful.
+
+### Download the original model
+First, clone the original model to some local directory:
+```console
+$ mkdir models && cd models
+$ git clone https://huggingface.co/user/model_name
+$ cd model_name
+$ git lfs install
+$ git lfs pull
+```
+
+### Set the MODEL_PATH
+The path to the downloaded model can be provided in two ways:
+
+**Option 1: Environment variable (recommended for iterative development)**
+```console
+export MODEL_PATH=~/work/ai/models/some_model
+```
+
+**Option 2: Command line argument (for one-off tasks)**
+```console
+make causal-convert-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+Command line arguments take precedence over environment variables when both are provided.
+
+In cases where the transformer implementation for the model has not been released
+yet it is possible to set the environment variable `UNRELEASED_MODEL_NAME` which
+will then cause the transformer implementation to be loaded explicitely and not
+use AutoModelForCausalLM:
+```
+export UNRELEASED_MODEL_NAME=SomeNewModel
+```
+
+### Inspecting the original tensors
+```console
+# Using environment variable
+(venv) $ make causal-inspect-original-model
+
+# Or using command line argument
+(venv) $ make causal-inspect-original-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+### Running the original model
+This is mainly to verify that the original model works, and to compare the output
+from the converted model.
+```console
+# Using environment variable
+(venv) $ make causal-run-original-model
+
+# Or using command line argument
+(venv) $ make causal-run-original-model MODEL_PATH=~/work/ai/models/some_model
+```
+This command will save two files to the `data` directory, one is a binary file
+containing logits which will be used for comparison with the converted model
+later, and the other is a text file which allows for manual visual inspection.
+
+### Model conversion
+After updates have been made to [gguf-py](../../gguf-py) to add support for the
+new model, the model can be converted to GGUF format using the following command:
+```console
+# Using environment variable
+(venv) $ make causal-convert-model
+
+# Or using command line argument
+(venv) $ make causal-convert-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+### Inspecting the converted model
+The converted model can be inspected using the following command:
+```console
+(venv) $ make inspect-converted-model
+```
+
+### Running the converted model
+```console
+(venv) $ make run-converted-model
+```
+
+### Model logits verfication
+The following target will run the original model and the converted model and
+compare the logits:
+```console
+(venv) $ make causal-verify-logits
+```
+
+### Quantizing the model
+The causal model can be quantized to GGUF format using the following command:
+```console
+(venv) $ make causal-quantize-Q8_0
+Quantized model saved to: /path/to/quantized/model-Q8_0.gguf
+Export the quantized model path to QUANTIZED_MODEL variable in your environment
+```
+This will show the path to the quantized model in the terminal, which can then
+be used to set the `QUANTIZED_MODEL` environment variable:
+```console
+export QUANTIZED_MODEL=/path/to/quantized/model-Q8_0.gguf
+```
+Then the quantized model can be run using the following command:
+```console
+(venv) $ make causal-run-quantized-model
+```
+
+
+## Embedding Language Model Conversion
+
+### Download the original model
+```console
+$ mkdir models && cd models
+$ git clone https://huggingface.co/user/model_name
+$ cd model_name
+$ git lfs install
+$ git lfs pull
+```
+
+The path to the embedding model can be provided in two ways:
+
+**Option 1: Environment variable (recommended for iterative development)**
+```console
+export EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+
+**Option 2: Command line argument (for one-off tasks)**
+```console
+make embedding-convert-model EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+
+Command line arguments take precedence over environment variables when both are provided.
+
+### Running the original model
+This is mainly to verify that the original model works and to compare the output
+with the output from the converted model.
+```console
+# Using environment variable
+(venv) $ make embedding-run-original-model
+
+# Or using command line argument
+(venv) $ make embedding-run-original-model EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+This command will save two files to the `data` directory, one is a binary
+file containing logits which will be used for comparison with the converted
+model, and the other is a text file which allows for manual visual inspection.
+
+### Model conversion
+After updates have been made to [gguf-py](../../gguf-py) to add support for the
+new model the model can be converted to GGUF format using the following command:
+```console
+(venv) $ make embedding-convert-model
+```
+
+### Run the converted model
+```console
+(venv) $ make embedding-run-converted-model
+```
+
+### Model logits verfication
+The following target will run the original model and the converted model (which
+was done manually in the previous steps) and compare the logits:
+```console
+(venv) $ make embedding-verify-logits
+```
+
+### llama-server verification
+To verify that the converted model works with llama-server, the following
+command can be used:
+```console
+(venv) $ make embedding-start-embedding-server
+```
+Then open another terminal and set the `EMBEDDINGS_MODEL_PATH` environment
+variable as this will not be inherited by the new terminal:
+```console
+(venv) $ make embedding-curl-embedding-endpoint
+```
+This will call the `embedding` endpoing and the output will be piped into
+the same verification script as used by the target `embedding-verify-logits`.
+
+The causal model can also be used to produce embeddings and this can be verified
+using the following commands:
+```console
+(venv) $ make causal-start-embedding-server
+```
+Then open another terminal and set the `MODEL_PATH` environment
+variable as this will not be inherited by the new terminal:
+```console
+(venv) $ make casual-curl-embedding-endpoint
+```
+
+### Quantizing the model
+The embedding model can be quantized to GGUF format using the following command:
+```console
+(venv) $ make embedding-quantize-Q8_0
+Quantized model saved to: /path/to/quantized/model-Q8_0.gguf
+Export the quantized model path to QUANTIZED_EMBEDDING_MODEL variable in your environment
+```
+This will show the path to the quantized model in the terminal, which can then
+be used to set the `QUANTIZED_EMBEDDING_MODEL` environment variable:
+```console
+export QUANTIZED_EMBEDDING_MODEL=/path/to/quantized/model-Q8_0.gguf
+```
+Then the quantized model can be run using the following command:
+```console
+(venv) $ make embedding-run-quantized-model
+```
+
+## Perplexity Evaluation
+
+### Simple perplexity evaluation
+This allows to run the perplexity evaluation without having to generate a
+token/logits file:
+```console
+(venv) $ make perplexity-run QUANTIZED_MODEL=~/path/to/quantized/model.gguf
+```
+This will use the wikitext dataset to run the perplexity evaluation and
+output the perplexity score to the terminal. This value can then be compared
+with the perplexity score of the unquantized model.
+
+### Full perplexity evaluation
+First use the converted, non-quantized, model to generate the perplexity evaluation
+dataset using the following command:
+```console
+$ make perplexity-data-gen CONVERTED_MODEL=~/path/to/converted/model.gguf
+```
+This will generate a file in the `data` directory named after the model and with
+a `.kld` suffix which contains the tokens and the logits for the wikitext dataset.
+
+After the dataset has been generated, the perplexity evaluation can be run using
+the quantized model:
+```console
+$ make perplexity-run-full QUANTIZED_MODEL=~/path/to/quantized/model-Qxx.gguf LOGITS_FILE=data/model.gguf.ppl
+```
+
+> 📝 **Note:** The `LOGITS_FILE` is the file generated by the previous command
+> can be very large, so make sure you have enough disk space available.
+
+## HuggingFace utilities
+The following targets are useful for creating collections and model repositories
+on Hugging Face in the the ggml-org. These can be used when preparing a relase
+to script the process for new model releases.
+
+For the following targets a `HF_TOKEN` environment variable is required.
+
+> 📝 **Note:** Don't forget to logout from Hugging Face after running these
+> commands, otherwise you might have issues pulling/cloning repositories as
+> the token will still be in use:
+> $ huggingface-cli logout
+> $ unset HF_TOKEN
+
+### Create a new Hugging Face Model (model repository)
+This will create a new model repsository on Hugging Face with the specified
+model name.
+```console
+(venv) $ make hf-create-model MODEL_NAME='TestModel' NAMESPACE="danbev"
+Repository ID:  danbev/TestModel-GGUF
+Repository created: https://huggingface.co/danbev/TestModel-GGUF
+```
+Note that we append a `-GGUF` suffix to the model name to ensure a consistent
+naming convention for GGUF models.
+
+### Upload a GGUF model to model repository
+The following target uploads a model to an existing Hugging Face model repository.
+```console
+(venv) $ make hf-upload-gguf-to-model MODEL_PATH=dummy-model1.gguf REPO_ID=danbev/TestModel-GGUF
+📤 Uploading dummy-model1.gguf to danbev/TestModel-GGUF/dummy-model1.gguf
+✅ Upload successful!
+🔗 File available at: https://huggingface.co/danbev/TestModel-GGUF/blob/main/dummy-model1.gguf
+```
+This command can also be used to update an existing model file in a repository.
+
+### Create a new Collection
+```console
+(venv) $ make hf-new-collection NAME=TestCollection DESCRIPTION="Collection for testing scripts" NAMESPACE=danbev
+🚀 Creating Hugging Face Collection
+Title: TestCollection
+Description: Collection for testing scripts
+Namespace: danbev
+Private: False
+✅ Authenticated as: danbev
+📚 Creating collection: 'TestCollection'...
+✅ Collection created successfully!
+📋 Collection slug: danbev/testcollection-68930fcf73eb3fc200b9956d
+🔗 Collection URL: https://huggingface.co/collections/danbev/testcollection-68930fcf73eb3fc200b9956d
+
+🎉 Collection created successfully!
+Use this slug to add models: danbev/testcollection-68930fcf73eb3fc200b9956d
+```
+
+### Add model to a Collection
+```console
+(venv) $ make hf-add-model-to-collection COLLECTION=danbev/testcollection-68930fcf73eb3fc200b9956d MODEL=danbev/TestModel-GGUF
+✅ Authenticated as: danbev
+🔍 Checking if model exists: danbev/TestModel-GGUF
+✅ Model found: danbev/TestModel-GGUF
+📚 Adding model to collection...
+✅ Model added to collection successfully!
+🔗 Collection URL: https://huggingface.co/collections/danbev/testcollection-68930fcf73eb3fc200b9956d
+
+🎉 Model added successfully!
+
+```
diff --git a/examples/model-conversion/logits.cpp b/examples/model-conversion/logits.cpp
new file mode 100644
index 0000000000000..2cac6a3b3eaae
--- /dev/null
+++ b/examples/model-conversion/logits.cpp
@@ -0,0 +1,209 @@
+#include "llama.h"
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <ctype.h>
+#include <filesystem>
+
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [prompt]\n", argv[0]);
+    printf("\n");
+}
+
+int main(int argc, char ** argv) {
+    std::string model_path;
+    std::string prompt = "Hello, my name is";
+    int ngl = 0;
+    bool embedding_mode = false;
+
+    {
+        int i = 1;
+        for (; i < argc; i++) {
+            if (strcmp(argv[i], "-m") == 0) {
+                if (i + 1 < argc) {
+                    model_path = argv[++i];
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-ngl") == 0) {
+                if (i + 1 < argc) {
+                    try {
+                        ngl = std::stoi(argv[++i]);
+                    } catch (...) {
+                        print_usage(argc, argv);
+                        return 1;
+                    }
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-embd-mode") == 0) {
+                if (i + 1 < argc) {
+                    try {
+                        embedding_mode = true;
+                    } catch (...) {
+                        print_usage(argc, argv);
+                        return 1;
+                    }
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else {
+                // prompt starts here
+                break;
+            }
+        }
+
+        if (model_path.empty()) {
+            print_usage(argc, argv);
+            return 1;
+        }
+
+        if (i < argc) {
+            prompt = argv[i++];
+            for (; i < argc; i++) {
+                prompt += " ";
+                prompt += argv[i];
+            }
+        }
+    }
+
+    ggml_backend_load_all();
+    llama_model_params model_params = llama_model_default_params();
+    model_params.n_gpu_layers = ngl;
+
+    llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // Extract basename from model_path
+    const char * basename = strrchr(model_path.c_str(), '/');
+    basename = (basename == NULL) ? model_path.c_str() : basename + 1;
+
+    char model_name[256];
+    strncpy(model_name, basename, 255);
+    model_name[255] = '\0';
+
+    char * dot = strrchr(model_name, '.');
+    if (dot != NULL && strcmp(dot, ".gguf") == 0) {
+        *dot = '\0';
+    }
+    printf("Model name: %s\n", model_name);
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+    const int n_prompt = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
+
+    std::vector<llama_token> prompt_tokens(n_prompt);
+    if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) {
+        fprintf(stderr, "%s: error: failed to tokenize the prompt\n", __func__);
+        return 1;
+    }
+
+    llama_context_params ctx_params = llama_context_default_params();
+    ctx_params.n_ctx = n_prompt;
+    ctx_params.n_batch = n_prompt;
+    ctx_params.no_perf = false;
+    if (embedding_mode) {
+        ctx_params.embeddings = true;
+        ctx_params.n_ubatch = ctx_params.n_batch;
+    }
+
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    printf("Input prompt: \"%s\"\n", prompt.c_str());
+    printf("Tokenized prompt (%d tokens): ", n_prompt);
+    for (auto id : prompt_tokens) {
+        char buf[128];
+        int n = llama_token_to_piece(vocab, id, buf, sizeof(buf), 0, true);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
+            return 1;
+        }
+        std::string s(buf, n);
+        printf("%s", s.c_str());
+    }
+    printf("\n");
+
+    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+
+    if (llama_decode(ctx, batch)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    float * logits;
+    int n_logits;
+    const char * type;
+
+    if (embedding_mode) {
+        logits = llama_get_embeddings(ctx);
+        n_logits = llama_model_n_embd(model) * batch.n_tokens;
+        type = "-embeddings";
+        printf("Embeddings size: %d\n", n_logits);
+    } else {
+        logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+        n_logits = llama_vocab_n_tokens(vocab);
+        type = "";
+        printf("Vocab size: %d\n", n_logits);
+    }
+
+    std::filesystem::create_directory("data");
+
+    // Save logits to binary file
+    char bin_filename[512];
+    snprintf(bin_filename, sizeof(bin_filename), "data/llamacpp-%s%s.bin", model_name, type);
+    printf("Saving logits to %s\n", bin_filename);
+
+    FILE * f = fopen(bin_filename, "wb");
+    if (f == NULL) {
+        fprintf(stderr, "%s: error: failed to open binary output file\n", __func__);
+        return 1;
+    }
+    fwrite(logits, sizeof(float), n_logits, f);
+    fclose(f);
+
+    // Also save as text for debugging
+    char txt_filename[512];
+    snprintf(txt_filename, sizeof(txt_filename), "data/llamacpp-%s%s.txt", model_name, type);
+    f = fopen(txt_filename, "w");
+    if (f == NULL) {
+        fprintf(stderr, "%s: error: failed to open text output file\n", __func__);
+        return 1;
+    }
+    for (int i = 0; i < n_logits; i++) {
+        fprintf(f, "%d: %.6f\n", i, logits[i]);  // Added index and changed format
+    }
+    fclose(f);
+
+    // Print first and last 10 logits for quick verification
+    printf("First 10 logits: ");
+    for (int i = 0; i < 10 && i < n_logits; i++) {
+        printf("%.6f ", logits[i]);
+    }
+    printf("\n");
+
+    printf("Last 10 logits: ");
+    for (int i = n_logits - 10; i < n_logits; i++) {
+        if (i >= 0) printf("%.6f ", logits[i]);
+    }
+    printf("\n\n");
+
+    printf("Logits saved to %s\n", bin_filename);
+    printf("Logits saved to %s\n", txt_filename);
+
+    llama_free(ctx);
+    llama_model_free(model);
+
+    return 0;
+}
diff --git a/examples/model-conversion/requirements.txt b/examples/model-conversion/requirements.txt
new file mode 100644
index 0000000000000..b8148b269a249
--- /dev/null
+++ b/examples/model-conversion/requirements.txt
@@ -0,0 +1,5 @@
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch~=2.6.0
+torchvision~=0.21.0
+transformers~=4.55.0
+huggingface-hub~=0.34.0
diff --git a/examples/model-conversion/scripts/causal/compare-embeddings-logits.sh b/examples/model-conversion/scripts/causal/compare-embeddings-logits.sh
new file mode 100755
index 0000000000000..287158f638717
--- /dev/null
+++ b/examples/model-conversion/scripts/causal/compare-embeddings-logits.sh
@@ -0,0 +1,43 @@
+#/bin/bash
+
+set -e
+
+MODEL_PATH="${1:-"$MODEL_PATH"}"
+MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+
+if [ -t 0 ]; then
+    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+else
+    # Process piped JSON data and convert to binary (matching logits.cpp format)
+    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)
+    python3 -c "
+import json
+import sys
+import struct
+
+data = json.load(sys.stdin)
+
+# Flatten all embeddings completely
+flattened = []
+for item in data:
+    embedding = item['embedding']
+    for token_embedding in embedding:
+        flattened.extend(token_embedding)
+
+print(f'Total embedding values: {len(flattened)}', file=sys.stderr)
+
+# Write as binary floats - matches logitc.cpp fwrite format
+with open('$TEMP_FILE', 'wb') as f:
+    for value in flattened:
+        f.write(struct.pack('f', value))
+"
+    CPP_EMBEDDINGS="$TEMP_FILE"
+    trap "rm -f $TEMP_FILE" EXIT
+fi
+
+python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
+    --cpp-embeddings $CPP_EMBEDDINGS \
+    --prompt "Hello world today" \
+    --causal
+
diff --git a/examples/model-conversion/scripts/causal/compare-logits.py b/examples/model-conversion/scripts/causal/compare-logits.py
new file mode 100755
index 0000000000000..fb959f0d566b4
--- /dev/null
+++ b/examples/model-conversion/scripts/causal/compare-logits.py
@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import sys
+import os
+from pathlib import Path
+
+def quick_logits_check(pytorch_file, llamacpp_file):
+    """Lightweight sanity check before NMSE"""
+
+    try:
+        pytorch_logits = np.fromfile(pytorch_file, dtype=np.float32)
+        llamacpp_logits = np.fromfile(llamacpp_file, dtype=np.float32)
+    except Exception as e:
+        print(f"❌ NOK: Failed to load files - {e}")
+        return False
+
+    # Check shapes match
+    if pytorch_logits.shape != llamacpp_logits.shape:
+        print(f"❌ NOK: Shape mismatch - PyTorch: {pytorch_logits.shape}, llama.cpp: {llamacpp_logits.shape}")
+        return False
+
+    # Calculate key metrics
+    diff = pytorch_logits - llamacpp_logits
+    abs_diff = np.abs(diff)
+    max_diff = np.max(abs_diff)
+
+    # Get top 10 predictions from both models
+    pytorch_top10 = np.argsort(pytorch_logits)[-10:][::-1]
+    llamacpp_top10 = np.argsort(llamacpp_logits)[-10:][::-1]
+    print(f"Top 10 PyTorch logits: {pytorch_logits[pytorch_top10]}")
+    print(f"Top 10 llama.cpp logits: {llamacpp_logits[llamacpp_top10]}")
+    print(f"Max absolute difference: {max_diff:.4f}")
+
+    if max_diff > 1.0:
+        print(f"❌ NOK: Large differences detected - max diff: {max_diff:.4f}")
+        return False
+
+    return True
+
+def main():
+    model_path = os.getenv('MODEL_PATH')
+    if not model_path:
+        print("Error: MODEL_PATH environment variable not set")
+        sys.exit(1)
+
+    if not os.path.exists(model_path):
+        print(f"Error: Model file not found: {model_path}")
+        sys.exit(1)
+
+    model_name = os.path.splitext(os.path.basename(model_path))[0]
+    data_dir = Path("data")
+
+    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
+    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    if not pytorch_file.exists():
+        print(f"Error: PyTorch logits file not found: {pytorch_file}")
+        print("Please run scripts/run-org-model.sh first to generate this file.")
+        sys.exit(1)
+
+    if not llamacpp_file.exists():
+        print(f"Error: llama.cpp logits file not found: {llamacpp_file}")
+        print("Please run scripts/run-converted-model.sh first to generate this file.")
+        sys.exit(1)
+
+    print("Checked all required files were found. Proceeding...\n")
+
+
+    print("🔍 GGML Model Validation for model ", model_name)
+    print("=" * 40)
+    print(f"PyTorch logits  : {pytorch_file}")
+    print(f"llama.cpp logits: {llamacpp_file}")
+    print()
+
+    success = quick_logits_check(pytorch_file, llamacpp_file)
+
+    # Exit with appropriate code
+    if success:
+        print("✅ OK: Lightweight model check successful!")
+        print("       Ok to proceed with NMSE check...")
+        sys.exit(0)
+    else:
+        print(f"❌ NOK: Top 10 predictions don't match - generation will differ")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/model-conversion/scripts/causal/convert-model.sh b/examples/model-conversion/scripts/causal/convert-model.sh
new file mode 100755
index 0000000000000..56b21f9baa673
--- /dev/null
+++ b/examples/model-conversion/scripts/causal/convert-model.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"
+OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
+TYPE="${OUTTYPE:-f16}"
+METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
+CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+
+echo "Model path: ${MODEL_PATH}"
+echo "Model name: ${MODEL_NAME}"
+echo "Data  type: ${TYPE}"
+echo "Converted model path:: ${CONVERTED_MODEL}"
+echo "Metadata override: ${METADATA_OVERRIDE}"
+python ../../convert_hf_to_gguf.py --verbose \
+    ${MODEL_PATH} \
+    --outfile ${CONVERTED_MODEL} \
+    --outtype ${TYPE} \
+    --metadata "${METADATA_OVERRIDE}"
+
+echo ""
+echo "The environment variable CONVERTED_MODEL can be set to this path using:"
+echo "export CONVERTED_MODEL=$(realpath ${CONVERTED_MODEL})"
diff --git a/examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.sh b/examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.sh
new file mode 100755
index 0000000000000..2fb54ab990e52
--- /dev/null
+++ b/examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.sh
@@ -0,0 +1,113 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import importlib
+import sys
+import torch
+import numpy as np
+
+from transformers import AutoTokenizer, AutoConfig, AutoModel, AutoModelForCausalLM
+from pathlib import Path
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+config = AutoConfig.from_pretrained(model_path)
+
+print("Model type:       ", config.model_type)
+print("Vocab size:       ", config.vocab_size)
+print("Hidden size:      ", config.hidden_size)
+print("Number of layers: ", config.num_hidden_layers)
+print("BOS token id:     ", config.bos_token_id)
+print("EOS token id:     ", config.eos_token_id)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}ForCausalLM"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+else:
+    model = AutoModelForCausalLM.from_pretrained(model_path)
+print(f"Model class: {type(model)}")
+#print(f"Model file: {type(model).__module__}")
+
+model_name = os.path.basename(model_path)
+print(f"Model name: {model_name}")
+
+prompt = "Hello world today"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+print(f"Input tokens: {input_ids}")
+print(f"Input text: {repr(prompt)}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+
+with torch.no_grad():
+    outputs = model(input_ids, output_hidden_states=True)
+
+    # Extract hidden states from the last layer
+    # outputs.hidden_states is a tuple of (num_layers + 1) tensors
+    # Index -1 gets the last layer, shape: [batch_size, seq_len, hidden_size]
+    last_hidden_states = outputs.hidden_states[-1]
+
+    # Get embeddings for all tokens
+    token_embeddings = last_hidden_states[0].cpu().numpy()  # Remove batch dimension
+
+    print(f"Hidden states shape: {last_hidden_states.shape}")
+    print(f"Token embeddings shape: {token_embeddings.shape}")
+    print(f"Hidden dimension: {token_embeddings.shape[-1]}")
+    print(f"Number of tokens: {token_embeddings.shape[0]}")
+
+    # Save raw token embeddings
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
+
+    # Save all token embeddings as binary
+    print(token_embeddings)
+    token_embeddings.astype(np.float32).tofile(bin_filename)
+
+    # Save as text for inspection
+    with open(txt_filename, "w") as f:
+        for i, embedding in enumerate(token_embeddings):
+            for j, val in enumerate(embedding):
+                f.write(f"{i} {j} {val:.6f}\n")
+
+    # Print embeddings per token in the requested format
+    print("\nToken embeddings:")
+    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
+    for i, embedding in enumerate(token_embeddings):
+        # Format: show first few values, ..., then last few values
+        if len(embedding) > 10:
+            # Show first 3 and last 3 values with ... in between
+            first_vals = " ".join(f"{val:8.6f}" for val in embedding[:3])
+            last_vals = " ".join(f"{val:8.6f}" for val in embedding[-3:])
+            print(f"embedding {i}: {first_vals}  ... {last_vals}")
+        else:
+            # If embedding is short, show all values
+            vals = " ".join(f"{val:8.6f}" for val in embedding)
+            print(f"embedding {i}: {vals}")
+
+    # Also show token info for reference
+    print(f"\nToken reference:")
+    for i, token in enumerate(tokens):
+        print(f"  Token {i}: {repr(token)}")
+
+    print(f"Saved bin logits to: {bin_filename}")
+    print(f"Saved txt logist to: {txt_filename}")
diff --git a/examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh b/examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh
new file mode 100755
index 0000000000000..64709f1798cb6
--- /dev/null
+++ b/examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m $CONVERTED_MODEL -embd-mode "Hello world today"
diff --git a/examples/model-conversion/scripts/causal/run-converted-model.sh b/examples/model-conversion/scripts/causal/run-converted-model.sh
new file mode 100755
index 0000000000000..e2762729e76a3
--- /dev/null
+++ b/examples/model-conversion/scripts/causal/run-converted-model.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" "Hello, my name is"
diff --git a/examples/model-conversion/scripts/causal/run-org-model.py b/examples/model-conversion/scripts/causal/run-org-model.py
new file mode 100755
index 0000000000000..f6188ea6f3382
--- /dev/null
+++ b/examples/model-conversion/scripts/causal/run-org-model.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import importlib
+from pathlib import Path
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+import torch
+import numpy as np
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+config = AutoConfig.from_pretrained(model_path)
+
+print("Model type:       ", config.model_type)
+print("Vocab size:       ", config.vocab_size)
+print("Hidden size:      ", config.hidden_size)
+print("Number of layers: ", config.num_hidden_layers)
+print("BOS token id:     ", config.bos_token_id)
+print("EOS token id:     ", config.eos_token_id)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+config = AutoConfig.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}ForCausalLM"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
+else:
+    model = AutoModelForCausalLM.from_pretrained(model_path)
+
+model_name = os.path.basename(model_path)
+# Printing the Model class to allow for easier debugging. This can be useful
+# when working with models that have not been publicly released yet and this
+# migth require that the concrete class is imported and used directly instead
+# of using AutoModelForCausalLM.
+print(f"Model class: {model.__class__.__name__}")
+
+prompt = "Hello, my name is"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+
+print(f"Input tokens: {input_ids}")
+print(f"Input text: {repr(prompt)}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+
+with torch.no_grad():
+    outputs = model(input_ids)
+    logits = outputs.logits
+
+    # Extract logits for the last token (next token prediction)
+    last_logits = logits[0, -1, :].cpu().numpy()
+
+    print(f"Logits shape: {logits.shape}")
+    print(f"Last token logits shape: {last_logits.shape}")
+    print(f"Vocab size: {len(last_logits)}")
+
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}.txt"
+
+    # Save to file for comparison
+    last_logits.astype(np.float32).tofile(bin_filename)
+
+    # Also save as text file for easy inspection
+    with open(txt_filename, "w") as f:
+        for i, logit in enumerate(last_logits):
+            f.write(f"{i}: {logit:.6f}\n")
+
+    # Print some sample logits for quick verification
+    print(f"First 10 logits: {last_logits[:10]}")
+    print(f"Last 10 logits: {last_logits[-10:]}")
+
+    # Show top 5 predicted tokens
+    top_indices = np.argsort(last_logits)[-5:][::-1]
+    print("Top 5 predictions:")
+    for idx in top_indices:
+        token = tokenizer.decode([idx])
+        print(f"  Token {idx} ({repr(token)}): {last_logits[idx]:.6f}")
+
+    print(f"Saved bin logits to: {bin_filename}")
+    print(f"Saved txt logist to: {txt_filename}")
diff --git a/examples/model-conversion/scripts/embedding/compare-embeddings-logits.sh b/examples/model-conversion/scripts/embedding/compare-embeddings-logits.sh
new file mode 100755
index 0000000000000..35b5d71984c6e
--- /dev/null
+++ b/examples/model-conversion/scripts/embedding/compare-embeddings-logits.sh
@@ -0,0 +1,42 @@
+#/bin/bash
+
+set -e
+
+MODEL_PATH="${1:-"$EMBEDDING_MODEL_PATH"}"
+MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+
+if [ -t 0 ]; then
+    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+else
+    # Process piped JSON data and convert to binary (matching logits.cpp format)
+    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)
+    python3 -c "
+import json
+import sys
+import struct
+
+data = json.load(sys.stdin)
+
+# Flatten all embeddings completely
+flattened = []
+for item in data:
+    embedding = item['embedding']
+    for token_embedding in embedding:
+        flattened.extend(token_embedding)
+
+print(f'Total embedding values: {len(flattened)}', file=sys.stderr)
+
+# Write as binary floats - matches logitc.cpp fwrite format
+with open('$TEMP_FILE', 'wb') as f:
+    for value in flattened:
+        f.write(struct.pack('f', value))
+"
+    CPP_EMBEDDINGS="$TEMP_FILE"
+    trap "rm -f $TEMP_FILE" EXIT
+fi
+
+python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
+    --cpp-embeddings $CPP_EMBEDDINGS \
+    --prompt "Hello world today"
+
diff --git a/examples/model-conversion/scripts/embedding/convert-model.sh b/examples/model-conversion/scripts/embedding/convert-model.sh
new file mode 100755
index 0000000000000..0609e35357bd4
--- /dev/null
+++ b/examples/model-conversion/scripts/embedding/convert-model.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+
+MODEL_NAME="${MODEL_NAME:-$(basename "$EMBEDDING_MODEL_PATH")}"
+OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
+TYPE="${OUTTYPE:-f16}"
+METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
+CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+
+echo "Model path: ${EMBEDDING_MODEL_PATH}"
+echo "Model name: ${MODEL_NAME}"
+echo "Data  type: ${TYPE}"
+echo "Converted model path:: ${CONVERTED_MODEL}"
+python ../../convert_hf_to_gguf.py --verbose \
+    ${EMBEDDING_MODEL_PATH} \
+    --outfile ${CONVERTED_MODEL} \
+    --outtype ${TYPE}
+
+echo ""
+echo "The environment variable CONVERTED_EMBEDDING MODEL can be set to this path using:"
+echo "export CONVERTED_EMBEDDING_MODEL=$(realpath ${CONVERTED_MODEL})"
diff --git a/examples/model-conversion/scripts/embedding/run-converted-model.sh b/examples/model-conversion/scripts/embedding/run-converted-model.sh
new file mode 100755
index 0000000000000..5896090411a79
--- /dev/null
+++ b/examples/model-conversion/scripts/embedding/run-converted-model.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_EMBEDDING_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_EMBEDDING_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "Hello world today"
diff --git a/examples/model-conversion/scripts/embedding/run-original-model.py b/examples/model-conversion/scripts/embedding/run-original-model.py
new file mode 100755
index 0000000000000..b9db0b893f13a
--- /dev/null
+++ b/examples/model-conversion/scripts/embedding/run-original-model.py
@@ -0,0 +1,116 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import numpy as np
+import importlib
+from pathlib import Path
+
+from transformers import AutoTokenizer, AutoConfig, AutoModel
+import torch
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or EMBEDDING_MODEL_PATH environment variable")
+
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}Model"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
+else:
+    model = AutoModel.from_pretrained(model_path)
+print(f"Model class: {type(model)}")
+#print(f"Model file: {type(model).__module__}")
+config = AutoConfig.from_pretrained(model_path)
+
+model_name = os.path.basename(model_path)
+
+texts = [ "Hello world today" ]
+
+encoded = tokenizer(
+    texts,
+    padding=True,
+    truncation=True,
+    return_tensors="pt"
+)
+
+tokens = encoded['input_ids'][0]
+token_strings = tokenizer.convert_ids_to_tokens(tokens)
+for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
+    print(f"{token_id:6d} -> '{token_str}'")
+
+with torch.no_grad():
+    outputs = model(**encoded)
+    hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+
+    # Extract embeddings for each token (matching LLAMA_POOLING_TYPE_NONE behavior)
+    all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
+
+    print(f"Hidden states shape: {hidden_states.shape}")
+    print(f"All embeddings shape: {all_embeddings.shape}")
+    print(f"Embedding dimension: {all_embeddings.shape[1]}")
+
+    # Print embeddings exactly like embedding.cpp does for LLAMA_POOLING_TYPE_NONE
+    n_embd = all_embeddings.shape[1]
+    n_embd_count = all_embeddings.shape[0]
+
+    print()  # Empty line to match C++ output
+
+    for j in range(n_embd_count):
+        embedding = all_embeddings[j]
+        print(f"embedding {j}: ", end="")
+
+        # Print first 3 values
+        for i in range(min(3, n_embd)):
+            print(f"{embedding[i]:9.6f} ", end="")
+
+        print(" ... ", end="")
+
+        # Print last 3 values
+        for i in range(n_embd - 3, n_embd):
+            print(f"{embedding[i]:9.6f} ", end="")
+
+        print()  # New line
+
+    print()  # Final empty line to match C++ output
+
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
+
+    # Save all embeddings flattened (matching what embedding.cpp would save if it did)
+    flattened_embeddings = all_embeddings.flatten()
+    flattened_embeddings.astype(np.float32).tofile(bin_filename)
+
+    with open(txt_filename, "w") as f:
+        f.write(f"# Model class: {model_name}\n")
+        f.write(f"# Tokens: {token_strings}\n")
+        f.write(f"# Shape: {all_embeddings.shape}\n")
+        f.write(f"# n_embd_count: {n_embd_count}, n_embd: {n_embd}\n\n")
+
+        for j in range(n_embd_count):
+            f.write(f"# Token {j} ({token_strings[j]}):\n")
+            for i, value in enumerate(all_embeddings[j]):
+                f.write(f"{j}_{i}: {value:.6f}\n")
+            f.write("\n")
+    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} tokens × {n_embd} dimensions)")
+    print("")
+    print(f"Saved bin embeddings to: {bin_filename}")
+    print(f"Saved txt embeddings to: {txt_filename}")
diff --git a/examples/model-conversion/scripts/readme.md.template b/examples/model-conversion/scripts/readme.md.template
new file mode 100644
index 0000000000000..87800a1b93a2f
--- /dev/null
+++ b/examples/model-conversion/scripts/readme.md.template
@@ -0,0 +1,13 @@
+---
+base_model:
+- {base_model}
+---
+# {model_name} GGUF
+
+Recommended way to run this model:
+
+```sh
+llama-server -hf {namespace}/{model_name}-GGUF -c 0 -fa
+```
+
+Then, access http://localhost:8080
diff --git a/examples/model-conversion/scripts/utils/check-nmse.py b/examples/model-conversion/scripts/utils/check-nmse.py
new file mode 100755
index 0000000000000..196a6210f048e
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/check-nmse.py
@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import sys
+import os
+import argparse
+from pathlib import Path
+
+def calculate_nmse(reference, test):
+    mse = np.mean((test - reference) ** 2)
+    ref_var = np.var(reference)
+    if ref_var == 0:
+        nmse = float('inf') if mse > 0 else 0.0
+        return mse, mse, ref_var
+
+    nmse = mse / ref_var
+
+    return nmse, mse, ref_var
+
+def load_logits(file_path):
+    if not os.path.exists(file_path):
+        raise FileNotFoundError(f"File not found: {file_path}")
+
+    if file_path.suffix == '.npy':
+        return np.load(file_path)
+    elif file_path.suffix == '.bin':
+        return np.fromfile(file_path, dtype=np.float32)
+    else:
+        # Try to load as text file
+        try:
+            # If it has index format "0: value", extract just values
+            data = []
+            with open(file_path, 'r') as f:
+                for line in f:
+                    if ':' in line:
+                        # Format: "index: value"
+                        value = float(line.split(':')[1].strip())
+                    else:
+                        # Just the value
+                        value = float(line.strip())
+                    data.append(value)
+            return np.array(data, dtype=np.float32)
+        except:
+            return np.loadtxt(file_path, dtype=np.float32)
+
+def interpret_nmse(nmse):
+    """Provide interpretation of NMSE value"""
+    if nmse == 0:
+        return "Perfect match", "🎉"
+    elif nmse < 1e-6:
+        return "Essentially identical", "✅"
+    elif nmse < 1e-4:
+        return "Excellent match", "✅"
+    elif nmse < 1e-3:
+        return "Very good match", "👍"
+    elif nmse < 1e-2:
+        return "Good match", "👍"
+    elif nmse < 0.1:
+        return "Acceptable match", "⚠️"
+    elif nmse < 1.0:
+        return "Poor match", "❌"
+    else:
+        return "Very poor match (worse than noise)", "❌"
+
+def main():
+    parser = argparse.ArgumentParser(description='Validate model logits')
+    parser.add_argument('-m', '--model-path', required=True,  help='Path to the model directory')
+    args = parser.parse_args()
+
+    model_name = os.path.splitext(os.path.basename(args.model_path))[0]
+    data_dir = Path("data")
+
+    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
+    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    print(f"Model name: {model_name}")
+    print(f"PyTorch logits file: {pytorch_file}")
+    print(f"llama.cpp logits file: {llamacpp_file}")
+
+    reference_file = pytorch_file
+    test_file = llamacpp_file
+
+    print("📊 NMSE Check for Model Comparison")
+    print("=" * 50)
+    print(f"Reference (ground truth): {reference_file}")
+    print(f"Test (to evaluate):       {test_file}")
+    print()
+
+    try:
+        print("Loading reference logits...")
+        reference = load_logits(reference_file)
+        print(f"  Shape: {reference.shape}, Type: {reference.dtype}")
+
+        print("Loading test logits...")
+        test = load_logits(test_file)
+        print(f"  Shape: {test.shape}, Type: {test.dtype}")
+
+        # Check shapes match
+        if reference.shape != test.shape:
+            print(f"\n❌ Error: Shape mismatch!")
+            print(f"  Reference: {reference.shape}")
+            print(f"  Test: {test.shape}")
+            sys.exit(1)
+
+        print(f"\n✅ Shapes match: {reference.shape}")
+
+        nmse, mse, ref_var = calculate_nmse(reference, test)
+
+        # Additional metrics
+        max_abs_error = np.max(np.abs(test - reference))
+        mean_abs_error = np.mean(np.abs(test - reference))
+
+        # Results
+        print(f"\n📈 METRICS")
+        print("=" * 30)
+        print(f"MSE (Mean Squared Error):     {mse:.6e}")
+        print(f"Reference Variance:           {ref_var:.6e}")
+        print(f"NMSE:                         {nmse:.6e}")
+        print(f"Max Absolute Error:           {max_abs_error:.6f}")
+        print(f"Mean Absolute Error:          {mean_abs_error:.6f}")
+
+        # NMSE in dB (common in signal processing)
+        if nmse > 0:
+            nmse_db = 10 * np.log10(nmse)
+            print(f"NMSE (dB):                    {nmse_db:.2f} dB")
+
+        # Interpretation
+        interpretation, emoji = interpret_nmse(nmse)
+        print(f"\n🎯 INTERPRETATION")
+        print("=" * 30)
+        print(f"{emoji} {interpretation}")
+
+        # Detailed guidance
+        print(f"\n📋 GUIDANCE")
+        print("=" * 30)
+        if nmse < 1e-3:
+            print("✅ EXCELLENT: Your GGML conversion is working very well!")
+            print("   The differences are negligible for practical use.")
+        elif nmse < 1e-2:
+            print("👍 GOOD: Your GGML conversion is working well.")
+            print("   Small differences are likely due to precision/quantization.")
+        elif nmse < 0.1:
+            print("⚠️  ACCEPTABLE: Conversion is working but with some differences.")
+            print("   Check if you're using quantization (Q4, Q8, etc.)")
+            print("   Test generation quality to see if it's acceptable.")
+        else:
+            print("❌ PROBLEMATIC: Large differences detected.")
+            print("   Check your conversion process for potential issues.")
+            print("   Verify you're using the same model weights.")
+
+        # NMSE benchmarks
+        print(f"\n📚 NMSE BENCHMARKS")
+        print("=" * 30)
+        print("< 1e-6:  Essentially identical")
+        print("< 1e-4:  Excellent (typical for good conversions)")
+        print("< 1e-3:  Very good")
+        print("< 1e-2:  Good (acceptable for most use cases)")
+        print("< 0.1:   Acceptable (may need verification)")
+        print("> 1.0:   Poor (worse than random)")
+
+        # Exit code based on NMSE
+        if nmse < 1e-2:
+            print(f"\n✅ RESULT: PASS (NMSE = {nmse:.2e})")
+            sys.exit(0)
+        else:
+            print(f"\n❌ RESULT: NEEDS REVIEW (NMSE = {nmse:.2e})")
+            sys.exit(1)
+
+    except Exception as e:
+        print(f"❌ Error: {e}")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/model-conversion/scripts/utils/create-collection-add-model.sh b/examples/model-conversion/scripts/utils/create-collection-add-model.sh
new file mode 100644
index 0000000000000..4809da6cb6fac
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/create-collection-add-model.sh
@@ -0,0 +1,6 @@
+
+COLLECTION_SLUG=$(python ./create_collection.py --return-slug)
+echo "Created collection: $COLLECTION_SLUG"
+
+# Use it in the next command
+python add_model_to_collection.py "$COLLECTION_SLUG" "username/my-model"
diff --git a/examples/model-conversion/scripts/utils/hf-add-model-to-collection.py b/examples/model-conversion/scripts/utils/hf-add-model-to-collection.py
new file mode 100755
index 0000000000000..7e38af3c136c6
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/hf-add-model-to-collection.py
@@ -0,0 +1,80 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import sys
+
+def add_model_to_collection(collection_slug, model_id, note=""):
+    """
+    Add a model to an existing collection
+
+    Args:
+        collection_slug: The slug of the collection (e.g., "username/collection-name-12345")
+        model_id: The model repository ID (e.g., "username/model-name")
+        note: Optional note about the model
+
+    Returns:
+        True if successful, False if failed
+    """
+
+    # Initialize API
+    api = HfApi()
+
+    try:
+        user_info = api.whoami()
+        print(f"✅ Authenticated as: {user_info['name']}")
+
+        # Verify the model exists
+        print(f"🔍 Checking if model exists: {model_id}")
+        try:
+            model_info = api.model_info(model_id)
+        except Exception as e:
+            print(f"❌ Model not found or not accessible: {model_id}")
+            print(f"Error: {e}")
+            return False
+
+        print(f"📚 Adding model to collection...")
+        api.add_collection_item(
+            collection_slug=collection_slug,
+            item_id=model_id,
+            item_type="model",
+            note=note
+        )
+
+        print(f"✅ Model added to collection successfully!")
+        print(f"🔗 Collection URL: https://huggingface.co/collections/{collection_slug}")
+
+        return True
+
+    except Exception as e:
+        print(f"❌ Error adding model to collection: {e}")
+        return False
+
+def main():
+    # This script requires that the environment variable HF_TOKEN is set with your
+    # Hugging Face API token.
+    api = HfApi()
+
+    parser = argparse.ArgumentParser(description='Add model to a Huggingface Collection')
+    parser.add_argument('--collection', '-c', help='The collection slug username/collection-hash', required=True)
+    parser.add_argument('--model', '-m', help='The model to add to the Collection', required=True)
+    parser.add_argument('--note', '-n', help='An optional note/description', required=False)
+    args = parser.parse_args()
+
+    collection = args.collection
+    model = args.model
+    note = args.note
+
+    success = add_model_to_collection(
+        collection_slug=collection,
+        model_id=model,
+        note=note
+    )
+
+    if success:
+        print("\n🎉 Model added successfully!")
+    else:
+        print("\n❌ Failed to add model to collection")
+        sys.exit(1)
+if __name__ == "__main__":
+    main()
diff --git a/examples/model-conversion/scripts/utils/hf-create-collection.py b/examples/model-conversion/scripts/utils/hf-create-collection.py
new file mode 100755
index 0000000000000..e0fa60af1ae7b
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/hf-create-collection.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import os
+import sys
+
+
+def create_collection(title, description, private=False, namespace=None, return_slug=False):
+    """
+    Create a new collection on Hugging Face
+
+    Args:
+        title: Collection title
+        description: Collection description
+        private: Whether the collection should be private (default: False)
+        namespace: Optional namespace (defaults to your username)
+
+    Returns:
+        Collection object if successful, None if failed
+    """
+
+    # Check if HF_TOKEN is available
+    token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
+    if not token:
+        print("❌ No HF_TOKEN or HUGGINGFACE_HUB_TOKEN found in environment variables")
+        print("Please set your Hugging Face token as an environment variable")
+        return None
+
+    # Initialize API
+    api = HfApi()
+
+    try:
+        # Test authentication first
+        user_info = api.whoami()
+        if not return_slug:
+            print(f"✅ Authenticated as: {user_info['name']}")
+
+        # Create the collection
+        if not return_slug:
+            print(f"📚 Creating collection: '{title}'...")
+        collection = api.create_collection(
+            title=title,
+            description=description,
+            private=private,
+            namespace=namespace
+        )
+
+        if not return_slug:
+            print(f"✅ Collection created successfully!")
+            print(f"📋 Collection slug: {collection.slug}")
+            print(f"🔗 Collection URL: https://huggingface.co/collections/{collection.slug}")
+
+        return collection
+
+    except Exception as e:
+        print(f"❌ Error creating collection: {e}")
+        return None
+
+def main():
+    # This script requires that the environment variable HF_TOKEN is set with your
+    # Hugging Face API token.
+    api = HfApi()
+
+    parser = argparse.ArgumentParser(description='Create a Huggingface Collection')
+    parser.add_argument('--name', '-n', help='The name/title of the Collection', required=True)
+    parser.add_argument('--description', '-d', help='The description for the Collection', required=True)
+    parser.add_argument('--namespace', '-ns', help='The namespace to add the Collection to', required=True)
+    parser.add_argument('--private', '-p', help='Create a private Collection', action='store_true')  # Fixed
+    parser.add_argument('--return-slug', '-s', help='Only output the collection slug', action='store_true')  # Fixed
+
+    args = parser.parse_args()
+
+    name = args.name
+    description = args.description
+    private = args.private
+    namespace = args.namespace
+    return_slug = args.return_slug
+
+    if not return_slug:
+        print("🚀 Creating Hugging Face Collection")
+        print(f"Title: {name}")
+        print(f"Description: {description}")
+        print(f"Namespace: {namespace}")
+        print(f"Private: {private}")
+
+    collection = create_collection(
+        title=name,
+        description=description,
+        private=private,
+        namespace=namespace,
+        return_slug=return_slug
+    )
+
+    if collection:
+        if return_slug:
+            print(collection.slug)
+        else:
+            print("\n🎉 Collection created successfully!")
+            print(f"Use this slug to add models: {collection.slug}")
+    else:
+        print("\n❌ Failed to create collection")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/model-conversion/scripts/utils/hf-create-model.py b/examples/model-conversion/scripts/utils/hf-create-model.py
new file mode 100755
index 0000000000000..09bb8511ef13e
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/hf-create-model.py
@@ -0,0 +1,63 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+
+# This script requires that the environment variable HF_TOKEN is set with your
+# Hugging Face API token.
+api = HfApi()
+
+def load_template_and_substitute(template_path, **kwargs):
+    try:
+        with open(template_path, 'r', encoding='utf-8') as f:
+            template_content = f.read()
+
+        return template_content.format(**kwargs)
+    except FileNotFoundError:
+        print(f"Template file '{template_path}' not found!")
+        return None
+    except KeyError as e:
+        print(f"Missing template variable: {e}")
+        return None
+
+parser = argparse.ArgumentParser(description='Create a new Hugging Face model repository')
+parser.add_argument('--model-name', '-m', help='Name for the model', required=True)
+parser.add_argument('--namespace', '-ns', help='Namespace to add the model to', required=True)
+parser.add_argument('--org-base-model', '-b', help='Original Base model name', default="")
+parser.add_argument('--no-card', action='store_true', help='Skip creating model card')
+parser.add_argument('--private', '-p', action='store_true', help='Create private model')
+
+args = parser.parse_args()
+
+repo_id = f"{args.namespace}/{args.model_name}-GGUF"
+print("Repository ID: ", repo_id)
+
+repo_url = api.create_repo(
+    repo_id=repo_id,
+    repo_type="model",
+    private=args.private,
+    exist_ok=False
+)
+
+if not args.no_card:
+    template_path = "scripts/readme.md.template"
+    model_card_content = load_template_and_substitute(
+        template_path,
+        model_name=args.model_name,
+        namespace=args.namespace,
+        base_model=args.org_base_model,
+    )
+
+    if model_card_content:
+        api.upload_file(
+            path_or_fileobj=model_card_content.encode('utf-8'),
+            path_in_repo="README.md",
+            repo_id=repo_id
+        )
+        print("Model card created successfully.")
+    else:
+        print("Failed to create model card.")
+
+print(f"Repository created: {repo_url}")
+
+
diff --git a/examples/model-conversion/scripts/utils/hf-upload-gguf-model.py b/examples/model-conversion/scripts/utils/hf-upload-gguf-model.py
new file mode 100755
index 0000000000000..15ccb1150e30b
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/hf-upload-gguf-model.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import os
+
+def upload_gguf_file(local_file_path, repo_id, filename_in_repo=None):
+    """
+    Upload a GGUF file to a Hugging Face model repository
+
+    Args:
+        local_file_path: Path to your local GGUF file
+        repo_id: Your repository ID (e.g., "username/model-name")
+        filename_in_repo: Optional custom name for the file in the repo
+    """
+
+    if not os.path.exists(local_file_path):
+        print(f"❌ File not found: {local_file_path}")
+        return False
+
+    if filename_in_repo is None:
+        filename_in_repo = os.path.basename(local_file_path)
+
+    if filename_in_repo is None or filename_in_repo == "":
+        filename_in_repo = os.path.basename(local_file_path)
+
+    print(f"📤 Uploading {local_file_path} to {repo_id}/{filename_in_repo}")
+
+    api = HfApi()
+
+    try:
+        api.upload_file(
+            path_or_fileobj=local_file_path,
+            path_in_repo=filename_in_repo,
+            repo_id=repo_id,
+            repo_type="model",
+            commit_message=f"Upload {filename_in_repo}"
+        )
+
+        print("✅ Upload successful!")
+        print(f"🔗 File available at: https://huggingface.co/{repo_id}/blob/main/{filename_in_repo}")
+        return True
+
+    except Exception as e:
+        print(f"❌ Upload failed: {e}")
+        return False
+
+# This script requires that the environment variable HF_TOKEN is set with your
+# Hugging Face API token.
+api = HfApi()
+
+parser = argparse.ArgumentParser(description='Upload a GGUF model to a Huggingface model repository')
+parser.add_argument('--gguf-model-path', '-m', help='The GGUF model file to upload', required=True)
+parser.add_argument('--repo-id', '-r', help='The repository to upload to', required=True)
+parser.add_argument('--name', '-o', help='The name in the model repository', required=False)
+args = parser.parse_args()
+
+upload_gguf_file(args.gguf_model_path, args.repo_id, args.name)
diff --git a/examples/model-conversion/scripts/utils/inspect-converted-model.sh b/examples/model-conversion/scripts/utils/inspect-converted-model.sh
new file mode 100755
index 0000000000000..e5b9324542e73
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/inspect-converted-model.sh
@@ -0,0 +1,14 @@
+#!/bin/bash
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+../../gguf-py/gguf/scripts/gguf_dump.py $CONVERTED_MODEL
diff --git a/examples/model-conversion/scripts/utils/inspect-org-model.py b/examples/model-conversion/scripts/utils/inspect-org-model.py
new file mode 100755
index 0000000000000..bc6f45a5fb7d0
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/inspect-org-model.py
@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import json
+from safetensors import safe_open
+from collections import defaultdict
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+# Check if there's an index file (multi-file model)
+index_path = os.path.join(model_path, "model.safetensors.index.json")
+single_file_path = os.path.join(model_path, "model.safetensors")
+
+if os.path.exists(index_path):
+    # Multi-file model
+    print("Multi-file model detected")
+
+    with open(index_path, 'r') as f:
+        index_data = json.load(f)
+
+    # Get the weight map (tensor_name -> file_name)
+    weight_map = index_data.get("weight_map", {})
+
+    # Group tensors by file for efficient processing
+    file_tensors = defaultdict(list)
+    for tensor_name, file_name in weight_map.items():
+        file_tensors[file_name].append(tensor_name)
+
+    print("Tensors in model:")
+
+    # Process each shard file
+    for file_name, tensor_names in file_tensors.items():
+        file_path = os.path.join(model_path, file_name)
+        print(f"\n--- From {file_name} ---")
+
+        with safe_open(file_path, framework="pt") as f:
+            for tensor_name in sorted(tensor_names):
+                tensor = f.get_tensor(tensor_name)
+                print(f"- {tensor_name} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+
+elif os.path.exists(single_file_path):
+    # Single file model (original behavior)
+    print("Single-file model detected")
+
+    with safe_open(single_file_path, framework="pt") as f:
+        keys = f.keys()
+        print("Tensors in model:")
+        for key in sorted(keys):
+            tensor = f.get_tensor(key)
+            print(f"- {key} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+
+else:
+    print(f"Error: Neither 'model.safetensors.index.json' nor 'model.safetensors' found in {model_path}")
+    print("Available files:")
+    if os.path.exists(model_path):
+        for item in sorted(os.listdir(model_path)):
+            print(f"  {item}")
+    else:
+        print(f"  Directory {model_path} does not exist")
+    exit(1)
diff --git a/examples/model-conversion/scripts/utils/perplexity-gen.sh b/examples/model-conversion/scripts/utils/perplexity-gen.sh
new file mode 100755
index 0000000000000..3db0b3fd27ac3
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/perplexity-gen.sh
@@ -0,0 +1,35 @@
+#!/bin/bash
+
+set -e
+
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+# Check if data/wikitext-2-raw directory exists
+if [ ! -d "ppl/wikitext-2-raw" ]; then
+    echo "ppl/wikitext-2-raw directory does not exist. Downloading..." >&2
+    mkdir -p ppl
+    pushd ppl
+    ./../../../scripts/get-wikitext-2.sh
+    popd
+fi
+
+mkdir -p ppl
+OUTPUTFILE="ppl/$(basename $CONVERTED_MODEL).kld"
+echo "Model: $CONVERTED_MODEL"
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $CONVERTED_MODEL \
+    -f ppl/wikitext-2-raw/wiki.test.raw \
+    --kl-divergence-base $OUTPUTFILE
+
+echo "Generated logits in $OUTPUTFILE"
+
diff --git a/examples/model-conversion/scripts/utils/perplexity-run-simple.sh b/examples/model-conversion/scripts/utils/perplexity-run-simple.sh
new file mode 100755
index 0000000000000..69b3438f598b3
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/perplexity-run-simple.sh
@@ -0,0 +1,27 @@
+#!/bin/bash
+
+set -e
+
+QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+
+if [ -z "$QUANTIZED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. QUANTIZED_MODEL environment variable" >&2
+    exit 1
+fi
+
+# Check if data/wikitext-2-raw directory exists
+if [ ! -d "ppl/wikitext-2-raw" ]; then
+    echo "ppl/wikitext-2-raw directory does not exist. Downloading..." >&2
+    mkdir -p ppl
+    pushd ppl
+    ./../../../scripts/get-wikitext-2.sh
+    popd
+fi
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL -f ppl/wikitext-2-raw/wiki.test.raw
+
+
diff --git a/examples/model-conversion/scripts/utils/perplexity-run.sh b/examples/model-conversion/scripts/utils/perplexity-run.sh
new file mode 100755
index 0000000000000..3bce7c8472153
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/perplexity-run.sh
@@ -0,0 +1,28 @@
+#!/bin/bash
+
+set -e
+
+QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+LOGITS_FILE="${1:-"$LOGITS_FILE"}"
+
+if [ -z "$QUANTIZED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. QUANTIZED_MODEL environment variable" >&2
+    exit 1
+fi
+
+if [ ! -f ${LOGITS_FILE} ]; then
+    echo "Error: logits file '${LOGITS_FILE} was not found"
+    echo "Did you run the perplexity-gen.sh script?"
+    exit 1
+fi
+
+echo "Model: $QUANTIZED_MODEL"
+echo "Data file: $LOGITS_FILE"
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL \
+    --kl-divergence-base $LOGITS_FILE \
+    --kl-divergence
diff --git a/examples/model-conversion/scripts/utils/quantize.sh b/examples/model-conversion/scripts/utils/quantize.sh
new file mode 100755
index 0000000000000..bcb87757543cb
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/quantize.sh
@@ -0,0 +1,34 @@
+#!/bin/bash
+
+set -e
+
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+QUANTIZED_TYPE="${2:-"$QUANTIZED_TYPE"}"
+QUANTIZED_MODEL=$CONVERTED_MODEL
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+# Process the quantized model filename
+if [[ "$QUANTIZED_MODEL" == *.gguf ]]; then
+    # Remove .gguf suffix, add quantized type, then add .gguf back
+    BASE_NAME="${QUANTIZED_MODEL%.gguf}"
+    QUANTIZED_MODEL="${BASE_NAME}-${QUANTIZED_TYPE}.gguf"
+else
+    echo "Error: QUANTIZED_MODEL must end with .gguf extension" >&2
+    exit 1
+fi
+
+
+cmake --build ../../build --target llama-quantize -j8
+
+../../build/bin/llama-quantize $CONVERTED_MODEL $QUANTIZED_MODEL $QUANTIZED_TYPE
+
+echo "Quantized model saved to: $QUANTIZED_MODEL"
diff --git a/examples/model-conversion/scripts/utils/run-embedding-server.sh b/examples/model-conversion/scripts/utils/run-embedding-server.sh
new file mode 100755
index 0000000000000..828fc47069772
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/run-embedding-server.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+#
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-server
+
+../../build/bin/llama-server -m $CONVERTED_MODEL \
+    --embedding \
+    --pooling none
diff --git a/examples/model-conversion/scripts/utils/semantic_check.py b/examples/model-conversion/scripts/utils/semantic_check.py
new file mode 100644
index 0000000000000..d2110480974e7
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/semantic_check.py
@@ -0,0 +1,179 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import argparse
+import os
+import importlib
+
+from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, AutoModel
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+def cosine_similarity(a, b=None):
+    a = np.asarray(a)
+    if b is None:
+        b = a
+    else:
+        b = np.asarray(b)
+
+    if a.ndim == 1:
+        a = a.reshape(1, -1)
+    if b.ndim == 1:
+        b = b.reshape(1, -1)
+
+    a_norms = np.linalg.norm(a, axis=1, keepdims=True)
+    b_norms = np.linalg.norm(b, axis=1, keepdims=True)
+
+    a_norms = np.where(a_norms == 0, 1e-8, a_norms)
+    b_norms = np.where(b_norms == 0, 1e-8, b_norms)
+
+    a_normalized = a / a_norms
+    b_normalized = b / b_norms
+
+    # Compute cosine similarity
+    return np.dot(a_normalized, b_normalized.T)
+
+def load_embeddings_from_file(filename, n_tokens, n_embd):
+    embeddings = np.fromfile(filename, dtype=np.float32)
+    return embeddings.reshape(n_tokens, n_embd)
+
+def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
+    np.set_printoptions(suppress=True, precision=6)
+    print("pytorch embeddings:");
+    print(python_emb)
+    print("llama.cpp embeddings:");
+    print(cpp_emb)
+    print(f"\n=== Prompt: '{prompt}' ===")
+    print(f"Tokens: {tokens}")
+    print(f"Embeddings shape: Python {python_emb.shape}, llama.cpp {cpp_emb.shape}")
+
+    n_tokens = len(tokens)
+
+    # 1. Direct embedding comparison
+    print(f"\n1. Raw Embedding Magnitude Comparison:")
+    # Check if the distance of each token embedding from the origin and compare
+    # if the vectors are on the same "sphere". This does not tell us about
+    # direction (meaning of the token embedding), just magnitude.
+    for i in range(n_tokens):
+        py_mag = np.linalg.norm(python_emb[i]) # calculate standard euclidean norm for Python embeddings
+        cpp_mag = np.linalg.norm(cpp_emb[i])   # calculate standard euclidean norm for llama.cpp embeddings
+        ratio = py_mag / cpp_mag if cpp_mag > 0 else float('inf')
+        print(f"   Token {i} ({tokens[i]}): Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
+
+    # 2. Cosine similarity between tokens within each model
+    # Here we check the direction of token embeddings to see if the have the
+    # same meaning (similarity). This is done by calculating cosine similarity
+    # of a pair of token embeddings within each model.
+    print(f"\n2. Within-Model Token Similarities:")
+    print("   Python model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([python_emb[i]], [python_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+
+    print("   llama.cpp model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([cpp_emb[i]], [cpp_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+
+    # 3. Cross-model similarity (same token position)
+    print(f"\n3. Cross-Model Same-Token Similarities:")
+    for i in range(n_tokens):
+        sim = cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0]
+        print(f"   Token {i} ({tokens[i]}): {sim:.4f}")
+
+    # 4. Similarity matrix comparison
+    print(f"\n4. Similarity Matrix Differences:")
+    py_sim_matrix = cosine_similarity(python_emb)
+    cpp_sim_matrix = cosine_similarity(cpp_emb)
+    diff_matrix = np.abs(py_sim_matrix - cpp_sim_matrix)
+
+    print(f"   Max difference: {np.max(diff_matrix):.4f}")
+    print(f"   Mean difference: {np.mean(diff_matrix):.4f}")
+    print(f"   RMS difference: {np.sqrt(np.mean(diff_matrix**2)):.4f}")
+
+    return {
+        'cross_model_similarities': [cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0] for i in range(n_tokens)],
+        'similarity_matrix_diff': diff_matrix,
+        'max_diff': np.max(diff_matrix),
+        'mean_diff': np.mean(diff_matrix),
+        'rms_diff': np.sqrt(np.mean(diff_matrix**2))
+    }
+
+def main():
+    parser = argparse.ArgumentParser(description='Test semantic similarity between Python and llama.cpp embeddings')
+    parser.add_argument('--model-path', '-m', required=True, help='Path to the original Python model')
+    parser.add_argument('--python-embeddings', '-pe', help='Path to pytorch embeddings "logits" binary file')
+    parser.add_argument('--cpp-embeddings', '-ce', help='Path to llama.cpp embeddings "logits" binary file')
+    parser.add_argument('--causal', '-c', default=False, help='if the model is causal (default: false)', action='store_true')
+    parser.add_argument('--prompt', '-p', default='Hello world today', help='Test prompt')
+
+    args = parser.parse_args()
+
+    print("Semantic Similarity Test Between Python and llama.cpp Embedding Models")
+    print("=" * 70)
+
+    # Single prompt detailed comparison
+    print(f"\nTesting with prompt: '{args.prompt}'")
+
+    # Load the python model to get configuration information and also to load the tokenizer.
+    print("Loading model and tokenizer using AutoTokenizer:", args.model_path)
+    tokenizer = AutoTokenizer.from_pretrained(args.model_path)
+    config = AutoConfig.from_pretrained(args.model_path)
+
+    if unreleased_model_name:
+        model_name_lower = unreleased_model_name.lower()
+        unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+        if args.causal:
+            class_name = f"{unreleased_model_name}ForCausalLM"
+        else:
+            class_name = f"{unreleased_model_name}Model"
+        print(f"Model class: {class_name}")
+        print(f"Importing unreleased model module: {unreleased_module_path}")
+
+        try:
+            model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+            model = model_class.from_pretrained(args.model_path)
+        except (ImportError, AttributeError) as e:
+            print(f"Failed to import or load model: {e}")
+            exit(1)
+    else:
+        if args.causal:
+            model = AutoModelForCausalLM.from_pretrained(args.model_path)
+        else:
+            model = AutoModel.from_pretrained(args.model_path)
+
+    encoded = tokenizer(args.prompt, return_tensors="pt")
+    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])
+    n_tokens = len(tokens)
+    print(f"n_tokens: {n_tokens}");
+    print(f"hidden_size: {model.config.hidden_size}")
+
+    # Load binary embeddings from data directory.
+    llamacpp_embeddings = load_embeddings_from_file(args.cpp_embeddings, n_tokens, model.config.hidden_size)
+    python_embeddings = load_embeddings_from_file(args.python_embeddings, n_tokens, model.config.hidden_size)
+
+    # Run comparison
+    results = test_single_prompt_similarity(python_embeddings, llamacpp_embeddings, tokens, args.prompt)
+
+    # Summary
+    print(f"\n=== SUMMARY ===")
+    avg_cross_sim = np.mean(results['cross_model_similarities'])
+    print(f"Average cross-model similarity: {avg_cross_sim:.4f}")
+    print(f"Similarity matrix RMS difference: {results['rms_diff']:.4f}")
+
+    # Quality assessment
+    if avg_cross_sim > 0.95:
+        print("✅ EXCELLENT: Models are highly similar")
+    elif avg_cross_sim > 0.90:
+        print("✅ VERY GOOD: Models are very similar")
+    elif avg_cross_sim > 0.80:
+        print("⚠️  GOOD: Models are reasonably similar")
+    elif avg_cross_sim > 0.70:
+        print("⚠️  FAIR: Models have some differences")
+    else:
+        print("❌ POOR: Models are significantly different")
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/passkey/README.md b/examples/passkey/README.md
index 2f19597c48d7f..cbaf28fd82f37 100644
--- a/examples/passkey/README.md
+++ b/examples/passkey/README.md
@@ -11,5 +11,5 @@ See the following PRs for more info:
 ### Usage
 
 ```bash
-make -j && ./llama-passkey -m ./models/llama-7b-v2/ggml-model-f16.gguf --junk 250
+llama-passkey -m ./models/llama-7b-v2/ggml-model-f16.gguf --junk 250
 ```
diff --git a/examples/retrieval/README.md b/examples/retrieval/README.md
index 6938a1e96ee35..51038cc36b1a8 100644
--- a/examples/retrieval/README.md
+++ b/examples/retrieval/README.md
@@ -15,7 +15,7 @@ https://github.com/ggml-org/llama.cpp/pull/6193
 `retrieval` example can be tested as follows:
 
 ```bash
-make -j && ./llama-retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
+llama-retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
 ```
 
 This chunks and embeds all given files and starts a loop requesting query inputs:
diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp
index 722cd7f40f088..a8e53f28eb597 100644
--- a/examples/speculative-simple/speculative-simple.cpp
+++ b/examples/speculative-simple/speculative-simple.cpp
@@ -59,6 +59,8 @@ int main(int argc, char ** argv) {
     }
 
     params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+    params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
+
     common_init_result llama_init_dft = common_init_from_params(params);
 
     //model_dft = llama_init_dft.model.get();
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index 0adffdb006bcf..8449406a6d27a 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -85,6 +85,8 @@ int main(int argc, char ** argv) {
     }
 
     params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+    params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
+
     common_init_result llama_init_dft = common_init_from_params(params);
 
     model_dft = llama_init_dft.model.get();
diff --git a/examples/sycl/win-build-sycl.bat b/examples/sycl/win-build-sycl.bat
index 6fc897b1486c8..862998e737569 100644
--- a/examples/sycl/win-build-sycl.bat
+++ b/examples/sycl/win-build-sycl.bat
@@ -18,8 +18,6 @@ if %errorlevel% neq 0 goto ERROR
 ::  for FP32
 cmake -G "Ninja" .. -DLLAMA_CURL=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
 if %errorlevel% neq 0 goto ERROR
-::  build example/main only
-::  make main
 
 ::  build all binary
 cmake --build . -j
diff --git a/examples/training/finetune.cpp b/examples/training/finetune.cpp
index 23bede49b1362..416d8d8f6c8f3 100644
--- a/examples/training/finetune.cpp
+++ b/examples/training/finetune.cpp
@@ -10,20 +10,20 @@
 #include <vector>
 
 #if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
+#pragma warning(disable: 4244 4267)  // possible loss of data
 #endif
 
 int main(int argc, char ** argv) {
     common_params params;
-
     params.escape = false;
 
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_FINETUNE)) {
         return 1;
     }
 
     if (params.use_mmap) {
-        LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n", __func__);
+        LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n",
+                __func__);
         params.use_mmap = false;
     }
     if (params.cache_type_k != GGML_TYPE_F32) {
@@ -38,11 +38,10 @@ int main(int argc, char ** argv) {
     common_init();
     llama_backend_init();
     llama_numa_init(params.numa);
-
     // load the model and apply lora adapter, if any
-    common_init_result llama_init = common_init_from_params(params);
-    llama_model_ptr   & model = llama_init.model;
-    llama_context_ptr & ctx   = llama_init.context;
+    common_init_result   llama_init = common_init_from_params(params);
+    llama_model_ptr    & model      = llama_init.model;
+    llama_context_ptr  & ctx        = llama_init.context;
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
@@ -55,31 +54,32 @@ int main(int argc, char ** argv) {
         LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
-    constexpr float val_split = 0.05f;
-
-    std::vector<llama_token> tokens = common_tokenize(ctx.get(), params.prompt, true);
-    ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get())/2);
-
-    struct ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr);
-    optimizer_params.adamw.alpha = 1e-7f; // learning rate
-
-    struct llama_opt_params lopt_params {
-        /*n_ctx_train     =*/ 0,
-        /*param_filter    =*/ llama_opt_param_filter_all,
-        /*param_filter_ud =*/ nullptr,
-        /*get_opt_pars    =*/ ggml_opt_get_constant_optimizer_params,
-        /*get_opt_pars_ud =*/ &optimizer_params,
+    std::vector<llama_token> tokens  = common_tokenize(ctx.get(), params.prompt, true);
+    ggml_opt_dataset_t       dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get()) / 2);
+
+    struct lr_opt & lr = params.lr;
+    LOG_INF("-optimizer %s -lr0 %.2g -wd %.2g -lr-min %.2g -min-epochs %.2g -epochs %d -period %.2g -val %.2g\n",
+            ggml_opt_optimizer_name(params.optimizer), (double) lr.lr0, (double) lr.wd, (double) lr.lr_min, (double) lr.decay_epochs,
+            (unsigned) lr.epochs, (double) params.n_batch / params.n_ubatch, (double) params.val_split);
+
+    struct llama_opt_params lopt_params{
+        /*n_ctx_train     =*/0,
+        /*param_filter    =*/llama_opt_param_filter_all,
+        /*param_filter_ud =*/nullptr,
+        /*get_opt_pars    =*/common_opt_lr_pars,
+        /*get_opt_pars_ud =*/&params.lr,
+        /*optimizer_type  =*/params.optimizer,
     };
     llama_opt_init(ctx.get(), model.get(), lopt_params);
 
-    const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split);
+    const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - params.val_split);
 
     ggml_opt_result_t result_train = ggml_opt_result_init();
     ggml_opt_result_t result_eval  = ggml_opt_result_init();
 
-    for (int epoch = 0; epoch < 2; ++epoch) {
+    for (lr.epoch = 0; lr.epoch < lr.epochs; ++lr.epoch) {
         llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
-            ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
+                        ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
         fprintf(stderr, "\n");
 
         ggml_opt_result_reset(result_train);
@@ -88,7 +88,7 @@ int main(int argc, char ** argv) {
     ggml_opt_result_free(result_train);
     ggml_opt_result_free(result_eval);
 
-    llama_model_save_to_file(model.get(), "finetuned-model.gguf");
+    llama_model_save_to_file(model.get(), params.out_file.c_str());
 
     llama_backend_free();
 
diff --git a/flake.nix b/flake.nix
index 0b5edf911fd06..bb02c8e52f9ad 100644
--- a/flake.nix
+++ b/flake.nix
@@ -36,9 +36,6 @@
   # ```
   # nixConfig = {
   #   extra-substituters = [
-  #     # Populated by the CI in ggml-org/llama.cpp
-  #     "https://llama-cpp.cachix.org"
-  #
   #     # A development cache for nixpkgs imported with `config.cudaSupport = true`.
   #     # Populated by https://hercules-ci.com/github/SomeoneSerge/nixpkgs-cuda-ci.
   #     # This lets one skip building e.g. the CUDA-enabled openmpi.
@@ -47,10 +44,8 @@
   #   ];
   #
   #   # Verify these are the same keys as published on
-  #   # - https://app.cachix.org/cache/llama-cpp
   #   # - https://app.cachix.org/cache/cuda-maintainers
   #   extra-trusted-public-keys = [
-  #     "llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc="
   #     "cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E="
   #   ];
   # };
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 1fb7abeaf088f..2ead001e2c610 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -158,7 +158,6 @@ option(GGML_CUDA                            "ggml: use CUDA"
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
@@ -188,6 +187,7 @@ option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_WEBGPU                          "ggml: use WebGPU"                                OFF)
 option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
+option(GGML_ZDNN                            "ggml: use zDNN"                                  OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
diff --git a/ggml/include/ggml-opt.h b/ggml/include/ggml-opt.h
index 74ec080a055ea..4703a05afe198 100644
--- a/ggml/include/ggml-opt.h
+++ b/ggml/include/ggml-opt.h
@@ -74,16 +74,26 @@ extern "C" {
         GGML_OPT_BUILD_TYPE_OPT     = 30,
     };
 
+    enum ggml_opt_optimizer_type {
+        GGML_OPT_OPTIMIZER_TYPE_ADAMW,
+        GGML_OPT_OPTIMIZER_TYPE_SGD,
+
+        GGML_OPT_OPTIMIZER_TYPE_COUNT
+    };
+
     // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
     struct ggml_opt_optimizer_params {
-        // AdamW optimizer parameters
         struct {
             float alpha; // learning rate
-            float beta1;
-            float beta2;
+            float beta1; // first AdamW momentum
+            float beta2; // second AdamW momentum
             float eps;   // epsilon for numerical stability
-            float wd;    // weight decay for AdamW, use 0.0f to disable
+            float wd;    // weight decay - 0.0f to disable
         } adamw;
+        struct {
+            float alpha; // learning rate
+            float wd;    // weight decay
+        } sgd;
     };
 
     // callback to calculate optimizer parameters prior to a backward pass
@@ -112,8 +122,11 @@ extern "C" {
 
         int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
 
-        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
-        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+        ggml_opt_get_optimizer_params get_opt_pars;    // callback for calculating optimizer parameters
+        void *                        get_opt_pars_ud; // userdata for calculating optimizer parameters
+
+        // only GGML_OPT_OPTIMIZER_TYPE_ADAMW needs m, v momenta per parameter tensor
+        enum ggml_opt_optimizer_type optimizer;
     };
 
     // get parameters for an optimization context with defaults set where possible
@@ -142,6 +155,10 @@ extern "C" {
     // get the gradient accumulator for a node from the forward graph
     GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
 
+    GGML_API enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t); //TODO consistent naming scheme
+
+    GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type);
+
     // ====== Optimization Result ======
 
     GGML_API ggml_opt_result_t ggml_opt_result_init(void);
@@ -226,12 +243,14 @@ extern "C" {
             struct ggml_tensor            * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
             ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
             enum ggml_opt_loss_type         loss_type,      // loss to minimize
+            enum ggml_opt_optimizer_type    optimizer,      // sgd or adamw
             ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
             int64_t                         nepoch,         // how many times the dataset should be iterated over
             int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
             float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
             bool                            silent);        // whether or not info prints to stderr should be suppressed
 
+
 #ifdef  __cplusplus
 }
 #endif
diff --git a/ggml/include/ggml-zdnn.h b/ggml/include/ggml-zdnn.h
new file mode 100644
index 0000000000000..c2c30c977c357
--- /dev/null
+++ b/ggml/include/ggml-zdnn.h
@@ -0,0 +1,16 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+GGML_BACKEND_API ggml_backend_t ggml_backend_zdnn_init(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_zdnn_reg(void);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 2f06e1e39b225..7e9c3c8c7a096 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -241,7 +241,16 @@
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24
 
+#define GGML_MROPE_SECTIONS   4
+
 #define GGML_UNUSED(x) (void)(x)
+#ifdef __CUDACC__
+template<typename... Args>
+__host__ __device__ constexpr inline void ggml_unused_vars_impl(Args&&...) noexcept {}
+#define GGML_UNUSED_VARS(...) ggml_unused_vars_impl(__VA_ARGS__)
+#else
+#define GGML_UNUSED_VARS(...) do { (void)sizeof((__VA_ARGS__, 0)); } while(0)
+#endif // __CUDACC__
 
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
@@ -503,6 +512,7 @@ extern "C" {
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
         GGML_OP_CONV_2D,
+        GGML_OP_CONV_3D,
         GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
@@ -540,6 +550,7 @@ extern "C" {
         GGML_OP_CROSS_ENTROPY_LOSS,
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
+        GGML_OP_OPT_STEP_SGD,
 
         GGML_OP_GLU,
 
@@ -1660,7 +1671,7 @@ extern "C" {
             struct ggml_tensor  * b,
             struct ggml_tensor  * c,
             int                   n_dims,
-            int                   sections[4],
+            int                   sections[GGML_MROPE_SECTIONS],
             int                   mode,
             int                   n_ctx_orig,
             float                 freq_base,
@@ -1686,6 +1697,22 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    GGML_API struct ggml_tensor * ggml_rope_multi_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[GGML_MROPE_SECTIONS],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
     GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1914,6 +1941,23 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
+    GGML_API struct ggml_tensor * ggml_conv_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // kernel [KW, KH, KD, IC * OC]
+            struct ggml_tensor  * b,   // input  [W, H, D, C * N]
+            int                   s0,  // stride
+            int                   s1,
+            int                   s2,
+            int                   p0,  // padding
+            int                   p1,
+            int                   p2,
+            int                   d0,  // dilation
+            int                   d1,
+            int                   d2,
+            int                   n_channels,
+            int                   n_batch,
+            int                   n_channels_out);
+
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,
@@ -2293,7 +2337,14 @@ extern "C" {
             struct ggml_tensor  * grad,
             struct ggml_tensor  * m,
             struct ggml_tensor  * v,
-            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
+            struct ggml_tensor  * adamw_params); // parameters such as the learning rate
+
+    // stochastic gradient descent step (with weight decay)
+    GGML_API struct ggml_tensor * ggml_opt_step_sgd(
+        struct ggml_context * ctx,
+        struct ggml_tensor *  a,
+        struct ggml_tensor *  grad,
+        struct ggml_tensor *  sgd_params); // alpha, weight decay
 
     //
     // automatic differentiation
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 177fb2821357f..2b5b8169d7513 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -382,6 +382,7 @@ ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
 ggml_add_backend(WebGPU)
+ggml_add_backend(zDNN)
 ggml_add_backend(OpenCL)
 
 foreach (target ggml-base ggml)
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 6c31513750c9b..5f02a710a14e0 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -49,6 +49,10 @@
 #include "ggml-webgpu.h"
 #endif
 
+#ifdef GGML_USE_ZDNN
+#include "ggml-zdnn.h"
+#endif
+
 #ifdef GGML_USE_OPENCL
 #include "ggml-opencl.h"
 #endif
@@ -180,6 +184,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_WEBGPU
         register_backend(ggml_backend_webgpu_reg());
 #endif
+#ifdef GGML_USE_ZDNN
+        register_backend(ggml_backend_zdnn_reg());
+#endif
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 1b9d29e911fcc..e34feccc98a5e 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -19,9 +19,8 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
-#include <string>
-#include <vector>
 #include <algorithm>
+#include <vector>
 
 #ifdef __APPLE__
 #include <sys/types.h>
@@ -1352,15 +1351,19 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
     struct ggml_backend_sched_split * splits = sched->splits;
 
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &splits[i];
+    ggml_tensor * prev_ids_tensor = nullptr;
+    std::vector<int32_t> ids;
+    std::vector<ggml_bitset_t> used_ids;
+
+    for (int split_id = 0; split_id < sched->n_splits; split_id++) {
+        struct ggml_backend_sched_split * split = &splits[split_id];
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
         // copy the input tensors to the split backend
-        for (int j = 0; j < split->n_inputs; j++) {
-            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
-            struct ggml_tensor * input = split->inputs[j];
+        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
+            struct ggml_tensor * input = split->inputs[input_id];
             struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
 
             if (input->flags & GGML_TENSOR_FLAG_INPUT) {
@@ -1378,16 +1381,104 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
-                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
-                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+
+                // when offloading MoE weights, we can reduce the amount of data copied by copying only the experts that are used
+                ggml_tensor * node = split->graph.nodes[0];
+                if (split->graph.n_nodes > 0 &&
+                    ggml_backend_buffer_get_usage(input->buffer) == GGML_BACKEND_BUFFER_USAGE_WEIGHTS &&
+                    ggml_backend_buffer_is_host(input->buffer) && (
+                    (node->src[0] == input_cpy && node->op == GGML_OP_MUL_MAT_ID)
+                    //|| (node->src[1] == input_cpy && node->op == GGML_OP_ADD_ID) /* GGML_OP_ADD_ID weights are small and not worth splitting */
+                    )) {
+
+                    const int64_t n_expert   = node->op == GGML_OP_MUL_MAT_ID ? input->ne[2] : input->ne[1];
+                    const size_t expert_size = node->op == GGML_OP_MUL_MAT_ID ? input->nb[2] : input->nb[1];
+
                     ggml_backend_synchronize(input_backend);
-                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                    } else {
-                        ggml_backend_synchronize(split_backend);
+
+                    // get the ids
+                    ggml_tensor * ids_tensor = node->src[2];
+                    ggml_backend_t ids_backend = split_backend;
+
+                    // if the ids tensor is also an input of the split, it may not have been copied yet to the split backend
+                    // in that case, we use the original ids tensor
+                    for (int i = input_id + 1; i < split->n_inputs; i++) {
+                        if (ids_tensor == tensor_copy(split->inputs[i], split_backend_id, sched->cur_copy)) {
+                            ids_tensor = split->inputs[i];
+                            ids_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[i]);
+                            break;
+                        }
+                    }
+
+                    if (ids_tensor != prev_ids_tensor) {
+                        ids.resize(ggml_nbytes(ids_tensor) / sizeof(int32_t));
+                        ggml_backend_tensor_get_async(ids_backend, ids_tensor, ids.data(), 0, ggml_nbytes(ids_tensor));
+                        ggml_backend_synchronize(ids_backend);
+
+                        // find the used experts
+                        used_ids.clear();
+                        used_ids.resize(ggml_bitset_size(n_expert));
+                        for (int64_t i1 = 0; i1 < ids_tensor->ne[1]; i1++) {
+                            for (int64_t i0 = 0; i0 < ids_tensor->ne[0]; i0++) {
+                                int32_t id = ids[i1 * ids_tensor->nb[1]/sizeof(int32_t) + i0 * ids_tensor->nb[0]/sizeof(int32_t)];
+                                GGML_ASSERT(id >= 0 && id < n_expert);
+                                ggml_bitset_set(used_ids.data(), id);
+                            }
+                        }
+
+                        prev_ids_tensor = ids_tensor;
+                    }
+
+                    // group consecutive experts and copy them together
+                    auto copy_experts = [&](int32_t first_id, int32_t last_id) {
+                        const size_t expert_offset = first_id * expert_size;
+                        const size_t expert_size_copy =  (last_id - first_id + 1) * expert_size;
+                        const size_t padding = std::min<size_t>(expert_size, 512);
+                        const size_t padding_end = last_id < n_expert - 1 ? padding : 0;
+
+                        ggml_backend_tensor_set_async(split_backend,
+                            input_cpy,
+                            (const uint8_t *)input->data + expert_offset, expert_offset,
+                            // copy a bit extra at the to ensure there are no NaNs in the padding of the last expert
+                            // this is necessary for MMQ in the CUDA backend
+                            expert_size_copy + padding_end);
+                    };
+
+                    int id = 0;
+                    while (!ggml_bitset_get(used_ids.data(), id)) {
+                        id++;
+                    }
+                    int32_t first_id = id;
+                    int32_t last_id = first_id;
+
+                    for (++id; id < n_expert; ++id) {
+                        if (!ggml_bitset_get(used_ids.data(), id)) {
+                            continue;
+                        }
+
+                        if (id == last_id + 1) {
+                            last_id = id;
+                            continue;
+                        }
+
+                        copy_experts(first_id, last_id);
+
+                        first_id = id;
+                        last_id = id;
+                    }
+                    copy_experts(first_id, last_id);
+                } else {
+                    // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                    // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                    if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                        ggml_backend_synchronize(input_backend);
+                        if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                            ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                        } else {
+                            ggml_backend_synchronize(split_backend);
+                        }
+                        ggml_backend_tensor_copy(input, input_cpy);
                     }
-                    ggml_backend_tensor_copy(input, input_cpy);
                 }
             }
         }
diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index 07d6b8b67d47c..8f65904b8fe51 100755
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -753,69 +753,55 @@ static void cann_copy(ggml_backend_cann_context& ctx, aclTensor* acl_src,
 void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];
 
-    aclTensor* acl_src = ggml_cann_create_tensor(src0);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
     if (ggml_are_same_shape(src0, dst)) {
+        aclTensor* acl_src = ggml_cann_create_tensor(src0);
+        aclTensor* acl_dst = ggml_cann_create_tensor(dst);
         if (dst->type == src0->type) {
             cann_copy(ctx, acl_src, acl_dst);
         } else {
             aclnn_cast(ctx, acl_src, acl_dst, ggml_cann_type_mapping(dst->type));
         }
+        ggml_cann_release_resources(ctx, acl_src, acl_dst);
     } else {
-        if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
-            if (dst->type == src0->type) {
-                size_t cpy_size = ggml_nbytes(dst);
-                ggml_cann_async_memcpy(ctx, dst->data, src0->data, cpy_size,
-                    ACL_MEMCPY_DEVICE_TO_DEVICE);
-                return;
-            } else {
-                ggml_cann_pool_alloc src_buffer_allocator(
-                    ctx.pool(),
-                    ggml_nelements(dst) * ggml_type_size(dst->type));
-                void* src_trans_buffer = src_buffer_allocator.get();
-                size_t src_trans_nb[GGML_MAX_DIMS];
-                src_trans_nb[0] = ggml_type_size(dst->type);
-                for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
-                }
-                aclTensor* src_trans_tensor = ggml_cann_create_tensor(
-                    src_trans_buffer, ggml_cann_type_mapping(dst->type),
-                    ggml_type_size(dst->type), src0->ne, src_trans_nb,
-                    GGML_MAX_DIMS);
-
-                aclnn_cast(ctx, acl_src, src_trans_tensor, ggml_cann_type_mapping(dst->type));
-                size_t cpy_size = ggml_nbytes(dst);
-                ggml_cann_async_memcpy(ctx, dst->data, src_trans_buffer, cpy_size,
-                    ACL_MEMCPY_DEVICE_TO_DEVICE);
-                ggml_cann_release_resources(ctx, src_trans_tensor);
-                return;
-            }
-        } else if (ggml_is_contiguous(dst)) {
-            ggml_cann_pool_alloc src_buffer_allocator(
-                ctx.pool(), ggml_nelements(dst) * ggml_type_size(dst->type));
-            void* src_trans_buffer = src_buffer_allocator.get();
+        void* src_trans_buffer = src0->data;
+        ggml_cann_pool_alloc src_buffer_allocator;
+        if (!ggml_is_contiguous(src0)) {
+            aclTensor* acl_src = ggml_cann_create_tensor(src0);
+            src_buffer_allocator.alloc(ctx.pool(),
+                ggml_nelements(src0) * ggml_type_size(src0->type));
+            src_trans_buffer = src_buffer_allocator.get();
             size_t src_trans_nb[GGML_MAX_DIMS];
-            src_trans_nb[0] = ggml_type_size(dst->type);
+            src_trans_nb[0] = ggml_type_size(src0->type);
             for (int i = 1; i < GGML_MAX_DIMS; i++) {
                 src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
             }
             aclTensor* src_trans_tensor = ggml_cann_create_tensor(
-                src_trans_buffer, ggml_cann_type_mapping(dst->type),
-                ggml_type_size(dst->type), src0->ne, src_trans_nb,
+                src_trans_buffer, ggml_cann_type_mapping(src0->type),
+                ggml_type_size(src0->type), src0->ne, src_trans_nb,
                 GGML_MAX_DIMS);
+            cann_copy(ctx, acl_src, src_trans_tensor);
+            ggml_cann_release_resources(ctx, acl_src, src_trans_tensor);
+        }
+
+        size_t src_reshape_nb[GGML_MAX_DIMS];
+        src_reshape_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            src_reshape_nb[i] = src_reshape_nb[i - 1] * dst->ne[i - 1];
+        }
 
-            aclnn_cast(ctx, acl_src, src_trans_tensor, ggml_cann_type_mapping(dst->type));
+        aclTensor* trans_acl_src = ggml_cann_create_tensor(src_trans_buffer,
+            ggml_cann_type_mapping(src0->type),ggml_type_size(src0->type),
+            dst->ne, src_reshape_nb, GGML_MAX_DIMS, ACL_FORMAT_ND);
+        aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-            size_t cpy_size = ggml_nbytes(dst);
-            ggml_cann_async_memcpy(ctx, dst->data, src_trans_buffer, cpy_size,
-                ACL_MEMCPY_DEVICE_TO_DEVICE);
-            ggml_cann_release_resources(ctx, src_trans_tensor);
-            return;
+        if (dst->type == src0->type) {
+            cann_copy(ctx, trans_acl_src, acl_dst);
         } else {
-            GGML_ABORT("Unsupport dst is not tontiguous.");
+            aclnn_cast(ctx, trans_acl_src, acl_dst, ggml_cann_type_mapping(dst->type));
         }
+        ggml_cann_release_resources(ctx, trans_acl_src, acl_dst);
     }
-    ggml_cann_release_resources(ctx, acl_src, acl_dst);
+    return;
 }
 
 /**
@@ -881,6 +867,86 @@ static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer,
     return acl_tensor;
 }
 
+/**
+ * @brief Fills a tensor with a scalar value.
+ *
+ * This function fills the destination tensor `acl_dst` with the scalar value
+ * `scalar`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param scalar The scalar value used to fill the tensor.
+ * @param acl_dst The destination tensor to be filled with the scalar value.
+ */
+static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
+                              aclTensor* acl_dst) {
+    auto acl_scalar = aclCreateScalar(&scalar, aclDataType::ACL_FLOAT);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst, acl_scalar);
+    ggml_cann_release_resources(ctx, acl_scalar);
+}
+
+/**
+ * @brief Get or expand a cached float32 tensor filled with a scalar value.
+ *
+ * This function manages cached device memory for float32 tensors. If the current
+ * cache size is insufficient for the requested tensor shape, the old memory will
+ * be released and new memory will be allocated. The allocated buffer is then
+ * initialized either with zeros (when @p value == 0.0f) or with the given scalar
+ * value using CANN operations. Finally, an aclTensor object is created from the
+ * cached memory and returned.
+ *
+ * @param ctx           The CANN backend context that manages device memory.
+ * @param buffer        A pointer to the cached device buffer (will be allocated
+ *                      or reallocated if necessary).
+ * @param cache_element The current number of cached elements. This will be
+ *                      updated when the cache is expanded.
+ * @param ne            The tensor shape array (number of elements in each dimension).
+ * @param nb            The stride size for each dimension.
+ * @param dims          The number of tensor dimensions.
+ * @param value         The scalar value used to fill the tensor (supports zero
+ *                      initialization via memset or arbitrary values via fill_scalar).
+ * @return              An aclTensor pointer created from the cached buffer.
+ */
+static aclTensor* get_f32_cache_acl_tensor(
+    ggml_backend_cann_context& ctx,
+    void** buffer,
+    int64_t &cache_element,
+    int64_t* ne,
+    size_t* nb,
+    int64_t dims,
+    float value) {
+    // Calculate total number of elements
+    int64_t n_element = 1;
+    for (int i = 0; i < dims; i++) {
+        n_element *= ne[i];
+    }
+    size_t size = n_element * sizeof(float);
+
+    // Allocate or expand cache if needed
+    if (cache_element < n_element) {
+        if (*buffer != nullptr) {
+            aclrtFree(*buffer);
+            *buffer = nullptr;
+        }
+
+        ACL_CHECK(aclrtMalloc(buffer, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        cache_element = n_element;
+
+        // Initialize cache
+        if (value == 0.0f) {
+            ACL_CHECK(aclrtMemsetAsync(*buffer, size, 0, size, ctx.stream()));
+        } else {
+            int64_t pool_ne[1] = { n_element };
+            size_t pool_nb[1] = { sizeof(float) };
+            aclTensor* acl_value = ggml_cann_create_tensor(
+                *buffer, ACL_FLOAT, sizeof(float), pool_ne, pool_nb, 1);
+            aclnn_fill_scalar(ctx, 1, acl_value);
+            ggml_cann_release_resources(ctx, acl_value);
+        }
+    }
+
+    return ggml_cann_create_tensor(*buffer, ACL_FLOAT, sizeof(float), ne, nb, dims);
+}
+
 void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src = dst->src[0];
 
@@ -889,20 +955,39 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     float eps;
     memcpy(&eps, dst->op_params, sizeof(float));
-    size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src);
-    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
-
-    aclTensor* acl_gamma = aclnn_values(
-        ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1,
-        ggml_cann_type_mapping(src->type), ggml_element_size(src));
-
-    size_t zero_tensor_n_bytes =
-        src->ne[1] * src->ne[2] * src->ne[3] * ggml_element_size(src);
-    ggml_cann_pool_alloc zero_tensor_allocator(ctx.pool(), zero_tensor_n_bytes);
-    aclTensor* acl_rstd =
-        aclnn_zero(ctx, zero_tensor_allocator.get(), zero_tensor_n_bytes,
-                   src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                   ggml_element_size(src));
+
+    // build gamma, one...
+    size_t acl_gamma_nb[GGML_MAX_DIMS];
+    acl_gamma_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        acl_gamma_nb[i] = acl_gamma_nb[i - 1] * src->ne[i - 1];
+    }
+    aclTensor* acl_gamma = get_f32_cache_acl_tensor(
+        ctx,
+        &ctx.f32_one_cache,
+        ctx.f32_one_cache_element,
+        src->ne,
+        acl_gamma_nb,
+        1,        // dims
+        1.0f      // value
+    );
+
+    // build rstd, zero...
+    size_t acl_rstd_nb[GGML_MAX_DIMS];
+    acl_rstd_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        acl_rstd_nb[i] = acl_rstd_nb[i - 1] * src->ne[i - 1];
+    }
+    aclTensor* acl_rstd = get_f32_cache_acl_tensor(
+        ctx,
+        &ctx.f32_zero_cache,
+        ctx.f32_zero_cache_element,
+        src->ne,
+        acl_rstd_nb,
+        GGML_MAX_DIMS,
+        0.0f      // value
+    );
+
     GGML_CANN_CALL_ACLNN_OP(ctx, RmsNorm, acl_src, acl_gamma, eps, acl_dst, acl_rstd);
     ggml_cann_release_resources(ctx, acl_src, acl_dst, acl_gamma, acl_rstd);
 }
@@ -917,14 +1002,13 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst,
 
     const int n_past = ((int32_t*)dst->op_params)[0];
 
-    size_t one_tensor_n_bytes = src->ne[0] * src->ne[1] * src->ne[2] *
-                                src->ne[3] * ggml_element_size(src);
-    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
+    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), ggml_nbytes(src));
+    void* buffer = one_tensor_allocator.get();
 
-    aclTensor* mask_tensor =
-        aclnn_values(ctx, one_tensor_allocator.get(), one_tensor_n_bytes,
-                     src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                     ggml_element_size(src), value);
+    aclTensor* mask_tensor = ggml_cann_create_tensor(buffer, ggml_cann_type_mapping(src->type),
+        ggml_type_size(src->type), src->ne, src->nb, GGML_MAX_DIMS);
+
+    aclnn_fill_scalar(ctx, value, mask_tensor);
 
     aclScalar* alpha = nullptr;
     float alphaValue = 1.0f;
@@ -1291,23 +1375,6 @@ void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
         tmp_permute_tensor, tmp_mul_tensor, acl_dst);
 }
 
-/**
- * @brief Fills a tensor with a scalar value.
- *
- * This function fills the destination tensor `acl_dst` with the scalar value
- * `scalar`.
- *
- * @param ctx The context for the CANN backend operations.
- * @param scalar The scalar value used to fill the tensor.
- * @param acl_dst The destination tensor to be filled with the scalar value.
- */
-static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
-                              aclTensor* acl_dst) {
-    auto acl_scalar = aclCreateScalar(&scalar, aclDataType::ACL_FLOAT);
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst, acl_scalar);
-    ggml_cann_release_resources(ctx, acl_scalar);
-}
-
 /**
  * @brief Raises each element of a tensor to the power of the corresponding
  * element in another tensor.
@@ -1330,160 +1397,196 @@ static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx,
 }
 
 /**
- * @brief   Applies the Alibi (Attention with Linear Biases) mechanism to the
- * @details This function implements the Alibi mechanism, which introduces
- *          learnable biases into the attention scores to simulate relative
- *          position encoding without the need for explicit positional
- *          embeddings.
- *
- * @param ctx          The backend CANN context for executing operations.
- * @param acl_src      The source tensor representing the query or key.
- * @param acl_position The position tensor containing relative positions.
- * @param acl_dst      The destination tensor where the result will be stored.
- * @param n_head       The number of attention heads.
- * @param src_ne       The dimensions of the source tensor.
- * @param src_nb0      The byte size of the first dimension of the source
- tensor.
- * @param max_bias     The maximum bias value used in the Alibi mechanism.
- * @param dst          The destination tensor object for additional metadata.
- *
- * The function performs the following steps:
- * 1. Calculates the logarithm floor of the number of heads to determine the
-      base for bias calculation.
- * 2. Initializes arrays with arithmetic sequences and fills them with bias
-      values.
- * 3. Computes the bias tensor based on the calculated biases and arithmetic
-      sequences.
- * 4. Reshapes the bias tensor to match the dimensions of the input tensors.
- * 5. Multiplies the position tensor by the bias tensor.
- * 6. Adds the result of the multiplication to the source tensor to produce the
-      final output.
+ * @brief Generate a range of values and apply a scalar base exponentiation.
+ *
+ * This function creates an evenly spaced sequence from `start` to `stop` (exclusive),
+ * with step size `step`, stores it in a temporary buffer, and then computes:
+ *
+ * @f[
+ * slope[i] = m^{\left( start + i \cdot step \right)}, \quad 0 \le i < size
+ * @f]
+ *
+ * The results are written to the provided @p slope_buffer.
+ *
+ * @param ctx           CANN backend context for memory allocation and operator execution.
+ * @param slope_buffer  Pointer to the output buffer (float array) for the computed slope values.
+ * @param m             Scalar base for the exponentiation.
+ * @param size          Number of elements in the generated sequence.
+ * @param start         Starting exponent offset.
+ * @param stop          Stopping exponent offset (exclusive).
+ * @param step          Step size for the exponent increment.
  */
-static void aclnn_alibi(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-                        aclTensor* acl_position, aclTensor* acl_dst,
-                        const int n_head, int64_t* src_ne, const size_t src_nb0,
-                        float max_bias, ggml_tensor* dst) {
-    const int64_t ne2_ne3 = src_ne[2] * src_ne[3];
-    GGML_ASSERT(src_nb0 == sizeof(float));
-    GGML_ASSERT(n_head == src_ne[2]);
-
-    const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
-
-    float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    // init arange
-    ggml_cann_pool_alloc arange_allocator(ctx.pool(),
-                                          ne2_ne3 * ggml_type_size(dst->type));
-    void* tmp_arange_buffer = arange_allocator.get();
+static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_buffer,
+    float m, int64_t size, float start, float stop, float step){
+    int64_t ne[] = {size};
+    size_t nb[] = {sizeof(float)};
 
-    // arange1: [1, ..., n_heads_log2_floor+1)
-    float start = 1;
-    float stop = n_heads_log2_floor + 1;
-    float step = 1;
-    int64_t n_elements_arange = n_heads_log2_floor;
+    ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(float));
+    void* arange_buffer = arange_allocator.get();
 
-    int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
-    size_t tmp_arange1_nb[] = {sizeof(dst->type)};
-    aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
-        tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_arange1_ne, tmp_arange1_nb,
-        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+    aclTensor* arange_tensor = ggml_cann_create_tensor(
+        arange_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
+    aclnn_arange(ctx, arange_tensor, start, stop, step, size);
 
-    aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
-
-    aclTensor* tmp_arange2_tensor = nullptr;
-    if (n_heads_log2_floor < ne2_ne3) {
-        // arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
-        start = 1;
-        stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
-        step = 2;
-        n_elements_arange = ne2_ne3 - n_heads_log2_floor;
-        int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
-        size_t tmp_arange2_nb[] = {sizeof(dst->type)};
-
-        aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
-            (char*)tmp_arange_buffer +
-                n_heads_log2_floor * ggml_type_size(dst->type),
-            ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
-            tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-        aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
-                     n_elements_arange);
-    }
+    aclTensor* slope_tensor = ggml_cann_create_tensor(
+        slope_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
 
-    // init mk_base
-    ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
-                                           ne2_ne3 * ggml_type_size(dst->type));
-    void* tmp_mk_base_buffer = mk_base_allocator.get();
-    int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
-    size_t tmp_mk_base1_nb[] = {sizeof(dst->type)};
-    aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
-        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_mk_base1_ne, tmp_mk_base1_nb,
-        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+    aclScalar* sc = aclCreateScalar(&m, aclDataType::ACL_FLOAT);
 
-    aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
-
-    aclTensor* tmp_mk_base2_tensor = nullptr;
-    if (n_heads_log2_floor < ne2_ne3) {
-        int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
-        size_t tmp_mk_base2_nb[] = {sizeof(dst->type)};
-        aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
-            (char*)tmp_mk_base_buffer +
-                n_heads_log2_floor * ggml_type_size(dst->type),
-            ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
-            tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-        aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
-    }
+    GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, sc, arange_tensor, slope_tensor);
+    ggml_cann_release_resources(ctx, sc, arange_tensor, slope_tensor);
+}
 
-    // init mk
-    int64_t tmp_mk_base_ne[] = {ne2_ne3};
-    size_t tmp_mk_base_nb[] = {sizeof(dst->type)};
-    aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
-        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_mk_base_ne, tmp_mk_base_nb,
-        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-    aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
-        tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_mk_base_ne, tmp_mk_base_nb,
-        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-    aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
+/**
+ * @brief Compute slope values for multiple attention heads based on ALiBi bias parameters.
+ *
+ * This function generates slope values for each attention head according to the ALiBi
+ * (Attention with Linear Biases) method. It splits the computation into two ranges depending
+ * on whether the head index is less than @p n_head_log2 or not, and uses different base values
+ * (`m0` and `m1`) for the exponentiation.
+ *
+ * @f[
+ * slope[h] =
+ * \begin{cases}
+ * m_0^{(h + 1)}, & h < n\_head\_log2 \\
+ * m_1^{\left( 2 \cdot (h - n\_head\_log2) + 1 \right)}, & h \geq n\_head\_log2
+ * \end{cases}
+ * \quad , \quad \text{if } max\_bias > 0
+ * @f]
+ *
+ * If @p max_bias <= 0, all slope values are set to 1.0.
+ *
+ * @param ctx           CANN backend context for memory allocation and operator execution.
+ * @param n_head        Total number of attention heads.
+ * @param slope_buffer  Pointer to the output buffer (float array) for storing slopes.
+ * @param max_bias      Maximum bias value for slope computation.
+ *
+*/
+static void aclnn_get_slope(ggml_backend_cann_context & ctx, int64_t n_head,
+    void* slope_buffer, float max_bias) {
+    const int n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+
+    float m0 = powf(2.0f, -(max_bias) / n_head_log2);
+    float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    // const float slope = (max_bias > 0.0f) ?
+    //                          h < n_head_log2 ?
+    //                              powf(m0, h + 1) :
+    //                              powf(m1, 2*(h - n_head_log2) + 1) :
+    //                          1.0f;
+    // arange1
+    float start = 0 + 1;
+    float end   = (n_head_log2 - 1) + 1;
+    float step  = 1;
+    float count = n_head_log2;
+    // end needs to be +1 because aclnn uses a left-closed, right-open interval.
+    aclnn_get_slope_inner(ctx, slope_buffer, m0, count, start, end + 1, step);
+    if (n_head_log2 < n_head) {
+        // arange2
+        start = 2 * (n_head_log2 - n_head_log2) + 1;
+        end   = 2 * ((n_head - 1) - n_head_log2) + 1;
+        step  = 2;
+        count = n_head - n_head_log2;
+        aclnn_get_slope_inner(
+            ctx, (char *) slope_buffer + n_head_log2 * sizeof(float),
+            m1, count, start, end + 1, step);
+    }
+}
 
-    // reshape mk
-    int64_t tmp_mk_ne[] = {1, 1, src_ne[2], src_ne[3]};
-    size_t tmp_mk_nb[GGML_MAX_DIMS];
-    tmp_mk_nb[0] = ggml_type_size(dst->type);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
+/**
+ * @brief Add ALiBi (Attention with Linear Biases) positional biases to the attention mask.
+ *
+ * This function computes the ALiBi slopes for each attention head (if max_bias > 0),
+ * multiplies them with the attention mask to produce bias tensors, and adds these biases
+ * to the destination tensor (@p dst).
+ *
+ * The function performs necessary broadcasting of the mask and slope tensors to match
+ * the shape of the destination tensor, then applies element-wise multiplication and addition
+ * using CANN operators.
+ *
+ * @param ctx         CANN backend context for memory management and operator execution.
+ * @param mask        Input attention mask tensor, assumed to be contiguous.
+ * @param dst         Destination tensor to which ALiBi biases will be added.
+ * @param dst_ptr     Pointer to the memory of the destination tensor.
+ * @param max_bias    Maximum bias value controlling the slope scaling.
+ *
+ * @note
+ * - Write data into dst_ptr using only the shape information of the dst tensor.
+ * - `GGML_MAX_DIMS + 2` is used to extend tensor dimensions for broadcasting.
+ */
+static void aclnn_add_alibi(ggml_backend_cann_context& ctx, ggml_tensor* mask,
+    ggml_tensor* dst, void* dst_ptr, float max_bias) {
+    void* slope_buffer = nullptr;
+    void* bias_buffer = nullptr;
+
+    if (max_bias > 0.0f) {
+        int64_t n_heads = dst->ne[2];
+        ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
+        slope_buffer = slope_allocator.get();
+        ggml_cann_pool_alloc bias_allocator(
+                    ctx.pool(), ggml_nelements(dst) * ggml_element_size(dst));
+        bias_buffer = bias_allocator.get();
+        aclnn_get_slope(ctx, n_heads, slope_buffer, max_bias);
     }
-    aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
-        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
-        ACL_FORMAT_ND);
 
-    // acl_position * mk
-    int64_t tmp_output_ne[] = {src_ne[0], src_ne[1], src_ne[2], src_ne[3]};
-    size_t tmp_output_nb[GGML_MAX_DIMS];
-    tmp_output_nb[0] = ggml_type_size(dst->type);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        tmp_output_nb[i] = tmp_output_nb[i - 1] * tmp_output_ne[i - 1];
+    // broadcast for mask, slop and dst;
+    int64_t nr2 = dst->ne[2] / mask->ne[2];
+    int64_t nr3 = dst->ne[3] / mask->ne[3];
+
+    // broadcast the mask across rows
+    int64_t mask_ne[] = { mask->ne[0], dst->ne[1], mask->ne[2], 1, mask->ne[3], 1 };
+    size_t  mask_nb[] = {
+        mask_nb[0] = mask->nb[0], mask_nb[1] = mask->nb[1], mask_nb[2] = mask->nb[2],
+        mask_nb[3] = mask->nb[2], mask_nb[4] = mask->nb[3], mask_nb[5] = mask->nb[3]
+    };
+
+    int64_t dst_ne[] = { dst->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], nr3 };
+    size_t  dst_nb[] = {
+        dst_nb[0] = dst->nb[0], dst_nb[1] = dst->nb[1], dst_nb[2] = dst->nb[2],
+        dst_nb[3] = dst->nb[2], dst_nb[4] = dst->nb[3], dst_nb[5] = dst->nb[3]
+    };
+
+    // slope is a 1 dim tensor, slope.ne2 == dst.ne2
+    int64_t slope_ne[] = { 1, 1, mask->ne[2], nr2, 1, 1 };
+    size_t  slope_nb[GGML_MAX_DIMS + 2];
+    slope_nb[0] = sizeof(float);
+    for (int i = 1; i < GGML_MAX_DIMS + 2; i++) {
+        slope_nb[i] = slope_nb[i - 1] * slope_ne[i - 1];
     }
-    ggml_cann_pool_alloc output_allocator(ctx.pool(), ggml_nbytes(dst));
-    void* tmp_output_buffer = output_allocator.get();
-    aclTensor* tmp_output_tensor = ggml_cann_create_tensor(
-        tmp_output_buffer, ggml_cann_type_mapping(dst->type),
-        ggml_type_size(dst->type), tmp_output_ne, tmp_output_nb, GGML_MAX_DIMS,
-        ACL_FORMAT_ND);
-    aclnn_mul(ctx, acl_position, tmp_mk_tensor, tmp_output_tensor);
 
-    // add
-    aclnn_add(ctx, tmp_output_tensor, acl_src, acl_dst);
-    ggml_cann_release_resources(ctx, tmp_arange1_tensor, tmp_arange2_tensor,
-        tmp_mk_base1_tensor, tmp_mk_base2_tensor, tmp_mk_base_tensor,
-        tmp_arange_tensor, tmp_mk_tensor, tmp_output_tensor);
+    aclTensor* acl_slope = ggml_cann_create_tensor(
+                            slope_buffer, ACL_FLOAT, sizeof(float),
+                            slope_ne, slope_nb, GGML_MAX_DIMS + 2);
+    aclTensor* acl_mask = ggml_cann_create_tensor(
+                            mask, mask_ne, mask_nb, GGML_MAX_DIMS + 2);
+
+    // write data into dst_ptr using only the shape information of the dst tensor.
+    aclTensor* acl_dst  = ggml_cann_create_tensor(
+                            dst_ptr, ggml_cann_type_mapping(dst->type),
+                            ggml_type_size(dst->type), dst_ne, dst_nb,
+                            GGML_MAX_DIMS + 2);
+
+    if (max_bias > 0.0f) {
+        int64_t bias_ne[] = { mask->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], 1 };
+        size_t  bias_nb[GGML_MAX_DIMS + 2];
+        bias_nb[0] = sizeof(float);
+        for (int i = 1; i < GGML_MAX_DIMS + 2; i++) {
+            bias_nb[i] = bias_nb[i - 1] * bias_ne[i - 1];
+        }
+        aclTensor* bias_tensor = ggml_cann_create_tensor(
+                                    bias_buffer, ACL_FLOAT, sizeof(float),
+                                    bias_ne, bias_nb, GGML_MAX_DIMS + 2);
+
+        aclnn_mul(ctx, acl_slope, acl_mask, bias_tensor);
+        aclnn_add(ctx, acl_dst, bias_tensor);
+        ggml_cann_release_resources(ctx, bias_tensor);
+    } else {
+        aclnn_add(ctx, acl_dst, acl_mask);
+    }
+    ggml_cann_release_resources(ctx, acl_slope, acl_mask, acl_dst);
 }
 
-void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_cann_dup(ctx, dst);
 }
 
@@ -1501,118 +1604,41 @@ void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
  * @param acl_dst The destination tensor where the softmax results will be
  * stored.
  */
-static void aclnn_softmax(ggml_backend_cann_context& ctx, aclTensor* acl_src,
-                          int64_t dim, aclTensor* acl_dst) {
+static void aclnn_softmax(ggml_backend_cann_context & ctx,
+    aclTensor* acl_src, int64_t dim, aclTensor * acl_dst) {
     GGML_CANN_CALL_ACLNN_OP(ctx, Softmax, acl_src, dim, acl_dst);
 }
 
-void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor* src0 = dst->src[0];
     ggml_tensor* src1 = dst->src[1];  // mask
 
     aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    aclTensor* acl_dst  = ggml_cann_create_tensor(dst);
 
-    float scale = 1.0f;
+    float scale    = 1.0f;
     float max_bias = 0.0f;
 
-    memcpy(&scale, (float*)dst->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (float*)dst->op_params + 1, sizeof(float));
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
 
     // input mul scale
     aclScalar* acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT);
+    ggml_cann_pool_alloc src_tensor_allocator(ctx.pool(), ggml_nbytes(src0));
+    void* src_tensor_buffer = src_tensor_allocator.get();
+    aclTensor* softmax_tensor = ggml_cann_create_tensor(
+        src_tensor_buffer, ggml_cann_type_mapping(src0->type),
+        ggml_element_size(src0), src0->ne, src0->nb,GGML_MAX_DIMS);
 
-    size_t n_bytes = ggml_nbytes(src0);
-    ggml_cann_pool_alloc mul_scale_allocator(ctx.pool(), n_bytes);
-    void* input_mul_scale_buffer = mul_scale_allocator.get();
-    aclTensor* acl_input_mul_scale_tensor = ggml_cann_create_tensor(
-        input_mul_scale_buffer, ACL_FLOAT, ggml_type_size(src0->type), src0->ne,
-        src0->nb, GGML_MAX_DIMS);
-
-    bool inplace = false;
-    aclnn_muls(ctx, acl_src0, scale, acl_input_mul_scale_tensor, inplace);
+    aclnn_muls(ctx, acl_src0, scale, softmax_tensor, false);
 
     // mask
-    aclTensor* acl_src1_fp32_tensor = nullptr;
-    aclTensor* tmp_mask_tensor = nullptr;
-    ggml_cann_pool_alloc src1_fp32_allocator(ctx.pool());
     if (src1) {
-        const bool use_f16 = src1->type == GGML_TYPE_F16;
-        if (use_f16) {
-            // cast to fp32
-            size_t n_bytes = ggml_nelements(src1) * sizeof(float_t);
-            size_t src1_fp32_nb[GGML_MAX_DIMS];
-            src1_fp32_nb[0] = sizeof(float_t);
-            for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                src1_fp32_nb[i] = src1_fp32_nb[i - 1] * src1->ne[i - 1];
-            }
-            src1_fp32_allocator.alloc(n_bytes);
-            void* src1_fp32_buffer = src1_fp32_allocator.get();
-            acl_src1_fp32_tensor = ggml_cann_create_tensor(
-                src1_fp32_buffer, ACL_FLOAT, sizeof(float), src1->ne,
-                src1_fp32_nb, GGML_MAX_DIMS);
-            aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
-            aclnn_cast(ctx, acl_src1, acl_src1_fp32_tensor, ACL_FLOAT);
-            ggml_cann_release_resources(ctx, acl_src1);
-        } else {
-            acl_src1_fp32_tensor = ggml_cann_create_tensor(src1);
-        }
-
-        // broadcast the mask across rows, only use ne11 of ne01 in mask
-        if (src1->ne[1] != src0->ne[1]) {
-            // mask shape: [1,1,ne11,ne10]
-            int64_t tmp_mask_ne[] = {src0->ne[0], src0->ne[1], 1, 1};
-            size_t tmp_mask_nb[GGML_MAX_DIMS];
-            tmp_mask_nb[0] = sizeof(float_t);
-            for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                tmp_mask_nb[i] = tmp_mask_nb[i - 1] * tmp_mask_ne[i - 1];
-            }
-            tmp_mask_tensor = ggml_cann_create_tensor(
-                src1->data, ACL_FLOAT, sizeof(float), tmp_mask_ne, tmp_mask_nb,
-                GGML_MAX_DIMS, ACL_FORMAT_ND);
-        }
-
-        // alibi
-        const int n_head = src0->ne[2];
-        const size_t src_nb0 = src0->nb[0];
-
-        n_bytes = ggml_nbytes(dst);
-        ggml_cann_pool_alloc output_allocator(ctx.pool(), n_bytes);
-        void* output_buffer = output_allocator.get();
-        aclTensor* alibi_output_tensor = ggml_cann_create_tensor(
-            output_buffer, ACL_FLOAT, ggml_type_size(dst->type), dst->ne,
-            dst->nb, GGML_MAX_DIMS);
-        if (max_bias <= 0.0f) {
-            // slope = 1.0
-            if (tmp_mask_tensor) {
-                aclnn_add(ctx, tmp_mask_tensor, acl_input_mul_scale_tensor,
-                          alibi_output_tensor);
-            } else {
-                aclnn_add(ctx, acl_src1_fp32_tensor, acl_input_mul_scale_tensor,
-                          alibi_output_tensor);
-            }
-        } else {
-            // slope != 1.0
-            if (tmp_mask_tensor) {
-                aclnn_alibi(ctx, acl_input_mul_scale_tensor, tmp_mask_tensor,
-                            alibi_output_tensor, n_head, src0->ne, src_nb0,
-                            max_bias, dst);
-            } else {
-                aclnn_alibi(ctx, acl_input_mul_scale_tensor,
-                            acl_src1_fp32_tensor, alibi_output_tensor, n_head,
-                            src0->ne, src_nb0, max_bias, dst);
-            }
-        }
-
-        // softmax
-        aclnn_softmax(ctx, alibi_output_tensor, 3, acl_dst);
-        ggml_cann_release_resources(ctx, alibi_output_tensor);
-    } else {
-        aclnn_softmax(ctx, acl_input_mul_scale_tensor, 3, acl_dst);
+        aclnn_add_alibi(ctx, src1, src0, src_tensor_buffer, max_bias);
     }
-
-    ggml_cann_release_resources(ctx, acl_src0, acl_src1_fp32_tensor, acl_dst,
-        acl_scale, acl_input_mul_scale_tensor, tmp_mask_tensor);
+    // softmax
+    aclnn_softmax(ctx, softmax_tensor, 3, acl_dst);
+    ggml_cann_release_resources(ctx, acl_src0, acl_dst, acl_scale, softmax_tensor);
 }
 
 /**
@@ -2209,86 +2235,129 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    // theta_scale arange, [0,1,...,ne00/2 - 1]
     int64_t theta_scale_length = ne00 / 2;
-    ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(),
-                                          theta_scale_length * sizeof(float_t));
-    void* theta_scale_buffer = theta_scale_allocator.get();
     int64_t theta_scale_ne[] = {theta_scale_length, 1, 1, 1};
     size_t theta_scale_nb[] = {sizeof(float_t), sizeof(float_t), sizeof(float_t),
                           theta_scale_length * sizeof(float_t)};
 
-    aclTensor* acl_theta_scale_tensor =
-        ggml_cann_create_tensor(theta_scale_buffer, ACL_FLOAT, sizeof(float_t),
-                                theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
-    float start = 0;
-    float step = 1;
-    float stop = ne00 / 2;
-    float n_elements = ne00 / 2;
-    aclnn_arange(ctx, acl_theta_scale_tensor, start, stop, step, n_elements);
-
-    // power
-    aclScalar* acl_theta_scale = aclCreateScalar(&theta_scale, aclDataType::ACL_FLOAT);
-    GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale, acl_theta_scale_tensor,
-                            acl_theta_scale_tensor);
-
-    // freq_scale
-    if (freq_scale != 1) {
-        aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
-    }
-
-    // freq_factors
-    if (src2) {
-        aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
-            src2->data, ggml_cann_type_mapping(src2->type),
-            ggml_type_size(src2->type), theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
-        aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor);
-        ggml_cann_release_resources(ctx, acl_freq_factors_tensor);
-    }
-
-    // position
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
     int64_t position_length = src1->ne[0];
     int64_t position_ne[] = {1, 1, position_length, 1};
     size_t position_nb[] = {sizeof(int32_t), sizeof(int32_t), sizeof(int32_t),
                             sizeof(int32_t) * position_length};
-    aclTensor* acl_position_tensor = ggml_cann_create_tensor(
-        src1->data, ggml_cann_type_mapping(src1->type),
-        ggml_type_size(src1->type), position_ne, position_nb, GGML_MAX_DIMS);
-
-    // power * position
-    int64_t theta_length = theta_scale_length * position_length;
-    ggml_cann_pool_alloc theta_allocator(ctx.pool(),
-                                         theta_length * sizeof(float_t));
-    void* theta_buffer = theta_allocator.get();
+
     int64_t theta_ne[] = {theta_scale_length, 1, position_length, 1};
     size_t theta_nb[GGML_MAX_DIMS];
     theta_nb[0] = sizeof(float_t);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
         theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
     }
-    aclTensor* acl_theta_tensor =
-        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
-                                theta_ne, theta_nb, GGML_MAX_DIMS);
-    aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
-              acl_theta_tensor);
-
-    // sin/cos
-    ggml_cann_pool_alloc sin_allocator(ctx.pool(),
-                                       theta_length * sizeof(float_t));
-    void* sin_buffer = sin_allocator.get();
-    aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
-        sin_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
-        GGML_MAX_DIMS, ACL_FORMAT_ND);
-    aclnn_sin(ctx, acl_theta_tensor, acl_sin_tensor);
 
-    ggml_cann_pool_alloc cos_allocator(ctx.pool(),
-                                       theta_length * sizeof(float_t));
-    void* cos_buffer = cos_allocator.get();
+    bool is_q = (std::strncmp(dst->name, "Qcur-", 5) == 0);
+    bool is_k = (std::strncmp(dst->name, "Kcur-", 5) == 0);
+
+    // used for accuracy testing
+    bool is_attention = is_q || is_k;
+
+    if(ctx.init_ptr == nullptr || !is_attention) {
+        // theta_scale arange, [0,1,...,ne00/2 - 1]
+        if(ctx.init_ptr != nullptr){
+            ACL_CHECK(aclrtFree(ctx.init_ptr));
+        }
+        ACL_CHECK(aclrtMalloc(&ctx.init_ptr, theta_scale_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+
+        aclTensor* acl_theta_scale_tensor =
+            ggml_cann_create_tensor(ctx.init_ptr, ACL_FLOAT, sizeof(float_t),
+                                    theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+        float start = 0;
+        float step = 1;
+        float stop = ne00 / 2;
+        float n_elements = ne00 / 2;
+        aclnn_arange(ctx, acl_theta_scale_tensor, start, stop, step, n_elements);
+
+        // power
+        aclScalar* acl_theta_scale = aclCreateScalar(&theta_scale, aclDataType::ACL_FLOAT);
+        GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale, acl_theta_scale_tensor,
+                                acl_theta_scale_tensor);
+
+        // freq_scale
+        if (freq_scale != 1) {
+            aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
+        }
+
+        // freq_factors
+        if (src2) {
+            aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
+                src2->data, ggml_cann_type_mapping(src2->type),
+                ggml_type_size(src2->type), theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+            aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor);
+            ggml_cann_release_resources(ctx, acl_freq_factors_tensor);
+        }
+        // release
+        ggml_cann_release_resources(ctx, acl_theta_scale_tensor,acl_theta_scale);
+    }
+
+    if(ctx.sin_ptr == nullptr) {
+        int64_t theta_length = theta_scale_length * ctx.max_prompt_length;
+        ACL_CHECK(aclrtMalloc(&ctx.sin_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+        ACL_CHECK(aclrtMalloc(&ctx.cos_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+    }
+    if(position_length > ctx.max_prompt_length) {
+        ctx.max_prompt_length = position_length;
+        int64_t theta_length = theta_scale_length * ctx.max_prompt_length;
+        ACL_CHECK(aclrtFree(ctx.sin_ptr));
+        ACL_CHECK(aclrtFree(ctx.cos_ptr));
+        ACL_CHECK(aclrtMalloc(&ctx.sin_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+        ACL_CHECK(aclrtMalloc(&ctx.cos_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+    }
+
+    bool is_fisrt_layer = (std::strncmp(dst->name, "Qcur-0", GGML_MAX_NAME) == 0);
+
+    if(is_fisrt_layer || !is_attention) {
+
+        aclTensor* acl_theta_scale_tensor =
+            ggml_cann_create_tensor(ctx.init_ptr, ACL_FLOAT, sizeof(float_t),
+                                    theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+
+        // position
+        aclTensor* acl_position_tensor = ggml_cann_create_tensor(
+            src1->data, ggml_cann_type_mapping(src1->type),
+            ggml_type_size(src1->type), position_ne, position_nb, GGML_MAX_DIMS);
+
+        // power * position
+        int64_t theta_length = theta_scale_length * position_length;
+        ggml_cann_pool_alloc theta_allocator(ctx.pool(),
+                                            theta_length * sizeof(float_t));
+        void* theta_buffer = theta_allocator.get();
+
+        aclTensor* acl_theta_tensor =
+            ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
+                                    theta_ne, theta_nb, GGML_MAX_DIMS);
+        aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
+                acl_theta_tensor);
+
+        // sin/cos
+        aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
+            ctx.sin_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
+            GGML_MAX_DIMS, ACL_FORMAT_ND);
+        aclnn_sin(ctx, acl_theta_tensor, acl_sin_tensor);
+
+        aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
+            ctx.cos_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
+            GGML_MAX_DIMS, ACL_FORMAT_ND);
+        aclnn_cos(ctx, acl_theta_tensor, acl_cos_tensor);
+
+        // release
+        ggml_cann_release_resources(ctx, acl_theta_scale_tensor, acl_position_tensor,
+            acl_theta_tensor, acl_sin_tensor, acl_cos_tensor);
+    }
+
+    aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
+            ctx.sin_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
+            GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
-        cos_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
-        GGML_MAX_DIMS, ACL_FORMAT_ND);
-    aclnn_cos(ctx, acl_theta_tensor, acl_cos_tensor);
+            ctx.cos_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
+            GGML_MAX_DIMS, ACL_FORMAT_ND);
 
     // attn_factor
     if (attn_factor != 1) {
@@ -2312,8 +2381,7 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     }
 
     // release
-    ggml_cann_release_resources(ctx, acl_theta_scale_tensor, acl_position_tensor,
-        acl_theta_tensor, acl_sin_tensor, acl_cos_tensor, acl_theta_scale);
+    ggml_cann_release_resources(ctx, acl_sin_tensor, acl_cos_tensor);
 }
 
 #ifdef __cplusplus
@@ -3208,104 +3276,24 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
             // Compute the slope if needed. Derived from ggml_cann_softmax().
             if(maxBias != 0.0f){
                 // alibi
-                const int64_t ne2_ne3 = src0->ne[2] * src0->ne[3];
-                const int64_t n_head = src0->ne[2];
-                const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
-                float m0 = powf(2.0f, -(maxBias) / n_heads_log2_floor);
-                float m1 = powf(2.0f, -(maxBias / 2.0f) / n_heads_log2_floor);
-                // init arange
-                ggml_cann_pool_alloc arange_allocator(ctx.pool(),
-                                                    ne2_ne3 * faElemSize);
-                void* tmp_arange_buffer = arange_allocator.get();
-
-                // arange1: [1, ..., n_heads_log2_floor+1)
-                float start = 1;
-                float stop = n_heads_log2_floor + 1;
-                float step = 1;
-                int64_t n_elements_arange = n_heads_log2_floor;
-
-                int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
-                size_t tmp_arange1_nb[] = {faElemSize};
-                aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
-                    tmp_arange_buffer, faDataType, faElemSize,
-                    tmp_arange1_ne, tmp_arange1_nb,
-                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-
-                aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
-
-                aclTensor* tmp_arange2_tensor = nullptr;
-                if (n_heads_log2_floor < ne2_ne3) {
-                    // arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
-                    start = 1;
-                    stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
-                    step = 2;
-                    n_elements_arange = ne2_ne3 - n_heads_log2_floor;
-                    int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
-                    size_t tmp_arange2_nb[] = {faElemSize};
-
-                    aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
-                        (char*)tmp_arange_buffer +
-                            n_heads_log2_floor * faElemSize,
-                        faDataType, faElemSize,
-                        tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-                    aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
-                                n_elements_arange);
+                const int64_t n_heads = src0->ne[2];
+                ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
+                void* slope_buffer = slope_allocator.get();
+                aclnn_get_slope(ctx, n_heads, slope_buffer, maxBias);
+
+                int64_t slope_ne[] = {1, 1, n_heads, 1};
+                size_t slope_nb[GGML_MAX_DIMS];
+                slope_nb[0] = sizeof(float);
+                for(int i = 1;i<GGML_MAX_DIMS;i++) {
+                    slope_nb[i] = slope_nb[i-1] * slope_ne[0];
                 }
 
-                // init mk_base
-                ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
-                                                    ne2_ne3 * faElemSize);
-                void* tmp_mk_base_buffer = mk_base_allocator.get();
-                int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
-                size_t tmp_mk_base1_nb[] = {faElemSize};
-                aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
-                    tmp_mk_base_buffer, faDataType, faElemSize,
-                    tmp_mk_base1_ne, tmp_mk_base1_nb,
-                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-
-                aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
-
-                aclTensor* tmp_mk_base2_tensor = nullptr;
-                if (n_heads_log2_floor < ne2_ne3) {
-                    int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
-                    size_t tmp_mk_base2_nb[] = {faElemSize};
-                    aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
-                        (char*)tmp_mk_base_buffer +
-                            n_heads_log2_floor * faElemSize,
-                        faDataType, faElemSize,
-                        tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-                    aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
-                }
+                aclTensor* slope_tensor = ggml_cann_create_tensor(
+                    slope_buffer, ACL_FLOAT, sizeof(float),
+                    slope_ne, slope_nb, GGML_MAX_DIMS);
+                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, slope_tensor);
 
-                // init mk
-                int64_t tmp_mk_base_ne[] = {ne2_ne3};
-                size_t tmp_mk_base_nb[] = {faElemSize};
-                aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
-                    tmp_mk_base_buffer, faDataType, faElemSize,
-                    tmp_mk_base_ne, tmp_mk_base_nb,
-                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-                aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
-                    tmp_arange_buffer, faDataType, faElemSize,
-                    tmp_mk_base_ne, tmp_mk_base_nb,
-                    GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
-                aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
-
-                // reshape mk
-                int64_t tmp_mk_ne[] = {1, 1, src0->ne[2], src0->ne[3]};
-                size_t tmp_mk_nb[GGML_MAX_DIMS];
-                tmp_mk_nb[0] = faElemSize;
-                for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
-                }
-                aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
-                    tmp_mk_base_buffer, faDataType, faElemSize,
-                    tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
-                    ACL_FORMAT_ND);
-                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, bcast_pse_tensor, tmp_mk_tensor);
-
-                ggml_cann_release_resources(ctx, tmp_arange1_tensor, tmp_arange2_tensor,
-                    tmp_mk_base1_tensor, tmp_mk_base2_tensor, tmp_mk_base_tensor,
-                    tmp_arange_tensor, tmp_mk_tensor);
+                ggml_cann_release_resources(ctx, slope_tensor);
             }
         }
 
diff --git a/ggml/src/ggml-cann/common.h b/ggml/src/ggml-cann/common.h
index 9d294f72b6779..5858bd3f6a197 100755
--- a/ggml/src/ggml-cann/common.h
+++ b/ggml/src/ggml-cann/common.h
@@ -368,6 +368,10 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+    void* init_ptr = nullptr;
+    void* sin_ptr = nullptr;
+    void* cos_ptr = nullptr;
+    int64_t max_prompt_length = 65536;
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
     std::unique_ptr<ggml_cann_graph> cann_graph;
@@ -375,6 +379,10 @@ struct ggml_backend_cann_context {
     cann_task_queue task_queue;
     bool async_mode;
     bool support_set_rows;
+    void* f32_zero_cache = nullptr;
+    void* f32_one_cache = nullptr;
+    int64_t f32_zero_cache_element = 0;
+    int64_t f32_one_cache_element = 0;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -414,6 +422,15 @@ struct ggml_backend_cann_context {
                 ACL_CHECK(aclrtDestroyStream(streams[i]));
             }
         }
+        if(init_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(init_ptr));
+        }
+        if(sin_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(sin_ptr));
+        }
+        if(cos_ptr != nullptr) {
+            ACL_CHECK(aclrtFree(cos_ptr));
+        }
     }
 
     /**
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index cf575b367500a..cb8af42ebf956 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2456,8 +2456,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             // value of paddingW should be at most half of kernelW
             return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
         }
-        case GGML_OP_SUM:
         case GGML_OP_DUP:
+        case GGML_OP_SUM:
         case GGML_OP_IM2COL:
         case GGML_OP_CONCAT:
         case GGML_OP_REPEAT:
@@ -2503,9 +2503,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (op->src[2]) {
                 return false;
             }
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
+            return true;
         case GGML_OP_FLASH_ATTN_EXT:{
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@@ -2532,11 +2530,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // DeepSeek MLA
                 return false;
             }
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-            if (op->src[0]->ne[3] != 1) {
-                return false;
-            }
             float logitSoftcap = 0.0f;
             memcpy(&logitSoftcap,  (float*)op->op_params + 2, sizeof(float));
             if(logitSoftcap != 0.0f) {
diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt
index f188d1638dc5d..ce0a3e1285eb0 100644
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -460,7 +460,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
             #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
             message(STATUS "z17 target")
-            list(APPEND ARCH_FLAGS -march=z17)
+            list(APPEND ARCH_FLAGS -march=arch15)
         else()
             message(STATUS "Unknown target")
             message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h
index b62e3158ee741..373408a9c0955 100644
--- a/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -40,18 +40,22 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
 // repack.cpp
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // repack.cpp
@@ -69,7 +73,6 @@
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -80,12 +83,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__loongarch64)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
@@ -103,12 +108,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__riscv)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
@@ -133,16 +140,16 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__s390x__)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -164,12 +171,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__wasm__)
 // quants.c
 #define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
@@ -195,10 +204,12 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #endif
diff --git a/ggml/src/ggml-cpu/arch/powerpc/quants.c b/ggml/src/ggml-cpu/arch/powerpc/quants.c
index 49aae7a23bba4..d3dfd049eaf14 100644
--- a/ggml/src/ggml-cpu/arch/powerpc/quants.c
+++ b/ggml/src/ggml-cpu/arch/powerpc/quants.c
@@ -278,6 +278,72 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_MXFP4 == 0);
+    static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+    const block_mxfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_MXFP4;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector unsigned char vshift4 = vec_splats((unsigned char)4);
+    vector float vsumf0 = vec_splats(0.0f);
+
+    vector signed char kv = vec_xl(0, (const signed char *)kvalues_mxfp4);
+
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d) *
+                                      GGML_E8M0_TO_FP32_HALF(x[ib].e));
+
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed char qxs = (vector signed char)vec_xl(0, x[ib].qs);
+
+        vector unsigned char lo_nibbles = (vector unsigned char)vec_and(qxs, lowMask);
+        vector unsigned char hi_nibbles = (vector unsigned char)vec_sr(qxs, vshift4);
+
+        vector signed char q4x0 = vec_perm(kv, kv, lo_nibbles);
+        vector signed char q4x1 = vec_perm(kv, kv, hi_nibbles);
+
+        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+
+        vector signed int vsumi0 = vec_splats((int32_t)0);
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi0 = vec_sum4s(qv1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vyd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+    sumf = vec_extract(vsumf0, 0);
+    *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
diff --git a/ggml/src/ggml-cpu/arch/s390/quants.c b/ggml/src/ggml-cpu/arch/s390/quants.c
index 7e4229d0e46a9..1c8176fb4d91f 100644
--- a/ggml/src/ggml-cpu/arch/s390/quants.c
+++ b/ggml/src/ggml-cpu/arch/s390/quants.c
@@ -23,6 +23,27 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(__VXE__) || defined(__VXE2__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const __attribute__((aligned(16))) uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b ) << 4
+static const __attribute__((aligned(16))) uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+
+// permute mask for byteswapping
+static const uint8x16_t v_kperm = (const uint8x16_t){
+     7,  6,  5,  4,  3,  2, 1, 0,
+    15, 14, 13, 12, 11, 10, 9, 8
+};
+#endif
+
 void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -241,6 +262,301 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float32x4_t v_sum0 = vec_splats(0.0f);
+    float32x4_t v_sum1 = vec_splats(0.0f);
+
+    uint32_t qh0, qh1;
+    uint64_t tmp0[4], tmp1[4];
+
+    const uint8x16_t v_m = vec_splats((uint8_t)0x0F);
+
+    #pragma GCC unroll 4
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q5_0 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
+
+        int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0));
+        int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2));
+        int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0));
+        int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2));
+
+        // required for fixing the byteorder
+        v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm);
+        v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm);
+        v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm);
+        v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm);
+
+        const uint8x16_t v_x0 = vec_xl(0, (const uint8_t *)x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, (const uint8_t *)x1->qs);
+
+        int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        const int8x16_t v_x0lf = vec_sub(v_x0l, v_qh0l);
+        const int8x16_t v_x0hf = vec_sub(v_x0h, v_qh0h);
+        const int8x16_t v_x1lf = vec_sub(v_x1l, v_qh1l);
+        const int8x16_t v_x1hf = vec_sub(v_x1h, v_qh1h);
+
+        const int8x16_t v_y0l = vec_xl(0,       (const int8_t *)y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_0/2, (const int8_t *)y0->qs);
+        const int8x16_t v_y1l = vec_xl(0,       (const int8_t *)y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_0/2, (const int8_t *)y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0);
+        v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
+    }
+
+    sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1);
+
+    #pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib];
+
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        uint64_t tmp[4];
+        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_1[(qh >> 24)       ];
+
+        int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0));
+        int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2));
+
+        // required for fixing the byteorder
+        v_qhl = vec_perm(v_qhl, v_qhl, v_kperm);
+        v_qhh = vec_perm(v_qhh, v_qhh, v_kperm);
+
+        const uint8x16_t v_x = vec_xl(0, (const uint8_t *)x0->qs);
+        int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        const int8x16_t v_xlf = vec_sub(v_xl, v_qhl);
+        const int8x16_t v_xhf = vec_sub(v_xh, v_qhh);
+
+        const int8x16_t v_yl = vec_xl(0,       (const int8_t *)y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, (const int8_t *)y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, vec_splats(0.0f));
+
+        sumf += vec_hsum(v_acc);
+    }
+
+    *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float32x4_t v_sum0 = vec_splats(0.0f);
+    float32x4_t v_sum1 = vec_splats(0.0f);
+
+    float summs0 = 0.0f;
+    float summs1 = 0.0f;
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    const uint8x16_t v_m = vec_splats((uint8_t)0x0F);
+
+    #pragma GCC unroll 4
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q5_1 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+        summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
+
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
+
+        int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0));
+        int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2));
+        int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0));
+        int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2));
+
+        // required for fixing the byteorder
+        v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm);
+        v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm);
+        v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm);
+        v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm);
+
+        const uint8x16_t v_x0 = vec_xl(0, x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, x1->qs);
+
+        const int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        const int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        const int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        const int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        const int8x16_t v_x0lf = vec_or(v_x0l, v_qh0l);
+        const int8x16_t v_x0hf = vec_or(v_x0h, v_qh0h);
+        const int8x16_t v_x1lf = vec_or(v_x1l, v_qh1l);
+        const int8x16_t v_x1hf = vec_or(v_x1h, v_qh1h);
+
+        const int8x16_t v_y0l = vec_xl(0      , y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_1/2, y0->qs);
+        const int8x16_t v_y1l = vec_xl(0      , y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_1/2, y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0);
+        v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
+    }
+
+    sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1) + summs0 + summs1;
+
+    #pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
+
+        float summs = GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        uint64_t tmp[4];
+        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_0[(qh >> 24)       ];
+
+        int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0));
+        int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2));
+
+        // required for fixing the byteorder
+        v_qhl = vec_perm(v_qhl, v_qhl, v_kperm);
+        v_qhh = vec_perm(v_qhh, v_qhh, v_kperm);
+
+        const uint8x16_t v_x = vec_xl(0, x0->qs);
+        const int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        const int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        const int8x16_t v_xlf = vec_or(v_xl, v_qhl);
+        const int8x16_t v_xhf = vec_or(v_xh, v_qhh);
+
+        const int8x16_t v_yl = vec_xl(0      , y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_1/2, y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, v_acc);
+
+        sumf += vec_hsum(v_acc) + summs;
+    }
+
+    *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp
index 37933a4b23fe8..d95bb6d8aafce 100644
--- a/ggml/src/ggml-cpu/arch/x86/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -511,38 +511,34 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
 #endif
 }
 
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+//
+// GEMV/GEMM templates
+//
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+
+// GEMV for 8x blocks of 32 4-bit quants with a single scale factor per block
+template<typename block_tx8>
+static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
+    static_assert(
+            std::is_same_v<block_tx8, block_q4_0x8> ||
+            std::is_same_v<block_tx8, block_iq4_nlx8>,
+            "Unsupported block type");
+
     const int qk = QK8_0;
     const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
-#if defined(__AVX2__)
-    // Lookup table to convert signed nibbles to signed bytes
-    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
-    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
     __m128i changemask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
     __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
 
     // Permute mask used for easier vector processing at later stages
     const __m256i m4b = _mm256_set1_epi8(0x0F);
 
-    int64_t b_nb = n / QK4_0;
+    int64_t b_nb = n / 32;
 
-    const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
+    const block_tx8  * b_ptr_start = (const block_tx8  *)vx;
     const block_q8_0 * a_ptr_start = (const block_q8_0 *)vy;
 
     // Process Q8_0 blocks one by one
@@ -551,17 +547,17 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         // Pointers to LHS blocks of block_q8_0 format
         const block_q8_0 * a_ptr = a_ptr_start + (y * nb);
 
-        // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
+        // Take group of eight blocks at each pass of the loop and perform dot product operation
         for (int64_t x = 0; x < nc / 8; x++) {
 
             // Pointers to RHS blocks
-            const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
 
             // Master FP accumulator
             __m256 acc_row = _mm256_setzero_ps();
 
             for (int64_t b = 0; b < nb; b++) {
-                // Load 8 blocks of Q4_0 interleaved as 8 bytes (B0 - B7)
+                // Load 8 blocks of 32 interleaved as 8 bytes (B0 - B7)
                 const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
                 const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 1);
                 const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 2);
@@ -578,8 +574,13 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 const __m256i rhs_vec_0123_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b)); // B0(24-31) B1(24-31) B2(24-31) B3(24-31)
                 const __m256i rhs_vec_4567_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b)); // B4(24-31) B5(24-31) B6(24-31) B7(24-31)
 
-                // Load the scale values for the 8 blocks interleaved in block_q4_0x8
-                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+                // Load the scale values for the 8 blocks interleaved in block_tx8
+                __m256 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+                }
 
                 // Load and convert to FP32 scale from block_q8_0
                 const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));
@@ -620,240 +621,782 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
             _mm256_storeu_ps(s + (y * nr + x * 8), acc_row);
         }
     }
-    return;
-
-#endif
-    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
-void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK_K;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+// GEMM for 8x blocks of 32 4-bit quants with a single scale factor per block
+template<typename block_tx8>
+static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
+    static_assert(
+            std::is_same_v<block_tx8, block_q4_0x8> ||
+            std::is_same_v<block_tx8, block_iq4_nlx8>,
+            "Unsupported block type");
 
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
+    const int qk = QK8_0;
+    const int nb = n / qk;
 
-#if defined(__AVX2__)
-    // Lookup table to convert signed nibbles to signed bytes
-    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
-    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
-    // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales
-    __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
-    __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0);
-    // Permute mask used for easier vector processing at later stages
-    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
+    const block_tx8    * b_ptr_start = (const block_tx8    *)vx;
+    const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
 
-    // Mask to extract nibbles from bytes
+    int64_t b_nb = n / 32;
+    int64_t y = 0;
+    // Mask to mask out nibbles from packed bytes
     const __m256i m4b = _mm256_set1_epi8(0x0F);
+    const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
+    // Permute mask used for easier vector processing at later stages
+    __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
+    int64_t xstart = 0;
+    int anr = nr - nr%16; // Used to align nr with boundary of 16
+#ifdef __AVX512F__
+    int anc = nc - nc%16; // Used to align nc with boundary of 16
+                          // Mask to mask out nibbles from packed bytes expanded to 512 bit length
+    const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
+    // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length
+    __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1);
 
-    int64_t b_nb = n / QK_K;
-
-    const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx;
-    const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
+    // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    for (; y < anr / 4; y += 4) {
 
-    // Process Q8_K blocks one by one
-    for (int64_t y = 0; y < nr; y++) {
+        const block_q8_0x4 * a_ptrs[4];
 
-        // Pointers to LHS blocks of block_q8_K format
-        const block_q8_K * a_ptr = a_ptr_start + (y * nb);
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
 
-        // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation
-        for (int64_t x = 0; x < nc / 8; x++) {
+        // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
 
-            // Pointers to RHS blocks
-            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+            const block_tx8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
+            const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
 
             // Master FP accumulators
-            __m256 acc_row = _mm256_setzero_ps();
-            __m256 acc_min_rows = _mm256_setzero_ps();
+            __m512 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
+            }
 
             for (int64_t b = 0; b < nb; b++) {
+                // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
+
+                const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
+                const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
+                const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
+                const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+
+                const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
 
-                // Load and convert to FP32 scale from block_q8_K
-                const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
-
-                // Load the scale values for the 8 blocks interleaved in block_q4_Kx8
-                // col_scale_f32 rearranged so as to multiply with appropriate quants
-                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
-                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
+                const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
 
-                __m256i iacc_b = _mm256_setzero_si256();
-                __m256i iacc_min_b = _mm256_setzero_si256();
+                const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
+                const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
 
-                const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums));
-                __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1)));
-                q8s = _mm256_permute2f128_si256(q8s, q8s, 0);
+                const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
+                const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
 
-                // Processes two sub blocks from each Q4_K in each iteration
-                for (int sb = 0; sb < QK_K / 64; sb++) {
+                const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
+                const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
 
-                    // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
-                    const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
-                    const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
-                    const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
-                    const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
-                    const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
-                    const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
-                    const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
+                // Shuffle pattern one - right side input
+                const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+                const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
 
-                    // 4-bit -> 8-bit
-                    // Values of the first sub block of eight block_q4_K structures for the sb loop
-                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b);
-                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b);
-                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b);
-                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b);
-                    const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b);
-                    const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b);
-                    const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b);
-                    const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b);
+                const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+                const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
 
-                    // Values of the second sub block of eight block_q4_K structures when sb = 1
-                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b);
-                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b);
-                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b);
-                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b);
-                    const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b);
-                    const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b);
-                    const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b);
-                    const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b);
+                const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+                const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
 
-                    uint32_t utmp_0[4], utmp_1[4];
+                const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+                const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
 
-                    // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together
-                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
-                    memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
-                    utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_0 = utmp_0[1] & kmask1;
-                    utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
-                    utmp_0[2] = uaux_0;
-                    utmp_0[0] &= kmask1;
+                // Shuffle pattern two - right side input
 
-                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
-                    memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
-                    utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_1 = utmp_1[1] & kmask1;
-                    utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
-                    utmp_1[2] = uaux_1;
-                    utmp_1[0] &= kmask1;
+                const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+                const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
 
-                    // Scales of first sub block in the sb loop
-                    const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
-                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask);
-                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
+                const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+                const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
 
-                    // Scales of second sub block in the sb loop
-                    __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
-                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask);
-                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
+                const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+                const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
 
-                    // Mins of first and second sub block of Q4_K block are arranged side by side
-                    __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
+                const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+                const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
 
-                    // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
-                    __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64)));
-                    __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64)));
-                    __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64)));
-                    __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64)));
+                // Scale values - Load the weight scale values of two block_tx8
+                __m512 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+                }
 
-                    lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0);
-                    lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0);
-                    lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0);
-                    lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
+                // Process LHS in pairs of rows
+                for (int rp = 0; rp < 4; rp++) {
 
-                    // Dot product done within 32 bit lanes and accumulated in the same vector
-                    // First done for first sub block and thenn for second sub block in each sb
-                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
-                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
-                    // ...........................................................................
-                    // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
+                    // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                    __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+                    __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
+                    __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
+                    __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+                    __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
+                    __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
+                    __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+                    __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
+                    __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
+                    __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+                    __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
+                    __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
 
+                    __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
+                    __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
+                    __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
+                    __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
+                    __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
+                    __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
+                    __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
+                    __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
 
-                    __m256i iacc_0 = _mm256_setzero_si256();
-                    __m256i iacc_1 = _mm256_setzero_si256();
+                    // Shuffle pattern one - left side input
 
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85)));
+                    const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                    const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
 
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255)));
+                    const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                    const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
 
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85)));
+                    const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                    const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
 
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255)));
+                    const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                    const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
 
-                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
+                    // Shuffle pattern two - left side input
 
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85)));
+                    const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                    const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
 
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255)));
+                    const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                    const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
 
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85)));
+                    const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                    const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
 
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255)));
+                    const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                    const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
 
-                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    // Resembles MMLAs into 2x2 matrices in ARM Version
+                    const __m512i zero = _mm512_setzero_epi32();
+                    __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
+                    __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
+                    __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
+                    __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
+                    __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
+                    __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
+                    __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
+                    __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
 
-                    // Accumulate the iacc value for one sb
-                    __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1);
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                    __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                    __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                    __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
 
-                    // Broadcast the bsums of the two sub blocks  of the iteration of Q8_K across the vector
-                    // Multiply-Add with corresponding mins of Q4_Kx8 with bsums
-                    __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0);
-                    __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01);
-                    q8s = _mm256_bsrli_epi128(q8s, 4);
 
-                    // Accumulate for the complete block
-                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
-                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+                    // Straighten out to make 4 row vectors
+                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                    const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
+                    const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+
+                    // Multiply with appropiate scales and accumulate
+                    acc_rows[rp * 4]     = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[rp * 4]);
+                    acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[rp * 4 + 1]);
+                    acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                    acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
                 }
+            }
 
-                // Multiply-Add with scale values for the complete super block
-                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
-                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
+        }
+    }
+
+    // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    for (; y < nr / 4; y ++) {
+        const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < anc / 8; x += 2) {
+
+            const block_tx8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
+            const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
 
+            // Master FP accumulators
+            __m512 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm512_setzero_ps();
             }
 
-            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
-            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
-            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+            for (int64_t b = 0; b < nb; b++) {
+                // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
+
+                const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
+                const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
+                const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
+                const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+
+                const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
+                const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
+
+                const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
+                const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
+
+                const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
+                const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
+
+                const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
+                const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
+
+                // Shuffle pattern one - right side input
+                const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+                const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+
+                const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+                const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+
+                const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+                const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+
+                const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+                const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+
+                // Shuffle pattern two - right side input
+
+                const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+                const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
+
+                const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+                const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+
+                const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+                const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+
+                const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+                const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+
+
+                // Scale values - Load the weight scale values of two block_tx8
+                __m512 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+                }
+
+                // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+                __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
+                __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
+                __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+                __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
+                __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
+                __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+                __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
+                __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
+                __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+                __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
+                __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+
+                __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
+                __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
+                __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
+                __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
+                __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
+                __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
+                __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
+                __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
+
+                // Shuffle pattern one - left side input
+
+                const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                // Shuffle pattern two - left side input
+
+                const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                // Resembles MMLAs into 2x2 matrices in ARM Version
+                const __m512i zero = _mm512_setzero_epi32();
+                __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
+                __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
+                __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
+                __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
+                __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
+                __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
+                __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
+                __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
+
+                // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+                // Straighten out to make 4 row vectors
+                __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+                __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+                __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+                __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+                // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
+                const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+
+                // Multiply with appropiate scales and accumulate
+                acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[0]);
+                acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[1]);
+                acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+            }
+
+            // Store the accumulated values
+            for (int i = 0; i < 4; i++) {
+                _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
         }
     }
+    if (anc != nc) {
+        xstart = anc/8;
+        y = 0;
+    }
+#endif // __AVX512F__
 
-#else
-    UNUSED(kmask1);
-    UNUSED(kmask2);
-    UNUSED(kmask3);
-    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+    // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+
+    for (; y < anr / 4; y += 4) {
+        const block_q8_0x4 * a_ptrs[4];
+
+        a_ptrs[0] = a_ptr_start + (y * nb);
+        for (int i = 0; i < 3; ++i) {
+            a_ptrs[i + 1] = a_ptrs[i] + nb;
+        }
+
+        // Take group of eight block_tx8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = xstart; x < nc / 8; x++) {
+
+            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_rows[16];
+            for (int i = 0; i < 16; i++) {
+                acc_rows[i] = _mm256_setzero_ps();
+            }
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+                const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+                const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+                const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+                const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+                const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+                const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+                const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+                // Shuffle pattern one - right side input
+                const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+                const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+                const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+                const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+                const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+                const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+                const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+                const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+                // Shuffle pattern two - right side input
+
+                const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+                const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+                const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+                const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+                const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+                const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+                const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+                const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+                // Scale values - Load the wight scale values of block_tx8
+                __m256 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                }
+
+                // Process LHS in groups of four
+                for (int rp = 0; rp < 4; rp++) {
+                    // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                    __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+                    __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+                    __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+                    __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+                    __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+                    __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+                    __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+                    __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+                    __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+                    __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+                    __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+                    __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+                    // Shuffle pattern one - left side input
+                    const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                    const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                    const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                    const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                    const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                    const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                    const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                    const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                    // Shuffle pattern two - left side input
+                    const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                    const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                    const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                    const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                    const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                    const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                    const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                    const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    // Resembles MMLAs into 2x2 matrices in ARM Version
+                    const __m256i zero = _mm256_setzero_si256();
+                    __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
+                    __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
+                    __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
+                    __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
+                    __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
+                    __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
+                    __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
+                    __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                    __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                    __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                    __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+                    // Straighten out to make 4 row vectors
+                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                    const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
+
+                    // Multiply with appropiate scales and accumulate
+                    acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+                    acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+                    acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                    acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32,  255)), acc_rows[rp * 4 + 3]);
+                }
+            }
+
+            // Store the accumulated values
+            for (int i = 0; i < 16; i++) {
+                _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
+        }
+    }
+
+    // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    for (; y < nr / 4; y ++) {
+        const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+        // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+        for (int64_t x = xstart; x < nc / 8; x++) {
+            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulators
+            __m256 acc_rows[4];
+            for (int i = 0; i < 4; i++) {
+                acc_rows[i] = _mm256_setzero_ps();
+            }
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b));  //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+                const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b));  //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+                const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b));  //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+                const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b));  //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+                const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b));  //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+                const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b));  //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+                const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b));  //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+                const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b));  //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+                // Shuffle pattern one - right side input
+                const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+                const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+                const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+                const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+                const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+                const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+                const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+                const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+                // Shuffle pattern two - right side input
+
+                const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+                const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+                const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+                const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+                const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+                const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+                const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+                const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+                // Scale values - Load the wight scale values of block_tx8
+                __m256 col_scale_f32;
+                if constexpr (
+                        std::is_same_v<block_tx8, block_q4_0x8> ||
+                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
+                    col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                }
+
+                // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+                __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+                __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+                __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+                __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+                __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+                __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+                __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+                __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+                __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+                __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+                __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+                // Shuffle pattern one - left side input
+
+                const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                // Shuffle pattern two - left side input
+
+                const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                // Resembles MMLAs into 2x2 matrices in ARM Version
+                const __m256i zero = _mm256_setzero_si256();
+                __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
+                __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
+                __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
+                __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
+                __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
+                __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
+                __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
+                __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+
+                // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+                // Straighten out to make 4 row vectors
+                __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
+
+                // Multiply with appropiate scales and accumulate
+                acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+                acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+                acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+            }
+
+            // Store the accumulated values
+            for (int i = 0; i < 4; i++) {
+                _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+            }
+        }
+    }
+}
+
+#endif // defined(__AVX2__) || defined(__AVX512F__)
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+    {
+        // Lookup table to convert signed nibbles to signed bytes
+        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+
+        gemv_q4_b32_8x8_q8_0_lut_avx<block_q4_0x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
+
+        return;
+    }
 #endif
+
+    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
-void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
     const int nb = n / qk;
     const int ncols_interleaved = 8;
     const int blocklen = 8;
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
 
     assert (n % qk == 0);
     assert (nc % ncols_interleaved == 0);
@@ -872,21 +1415,18 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     // Lookup table to convert signed nibbles to signed bytes
     __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
     signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
-    // Shuffle masks to rearrange delta values to multiply with appropriate scales
+    // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales
     __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+    __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0);
     // Permute mask used for easier vector processing at later stages
     __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
 
-    const __m256i m3b = _mm256_set1_epi8(3);
-    const __m128i m4b_sse = _mm_set1_epi8(0xF);
-
-    //Mask to get appropriate scales
-    __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
-    __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
+    // Mask to extract nibbles from bytes
+    const __m256i m4b = _mm256_set1_epi8(0x0F);
 
     int64_t b_nb = n / QK_K;
 
-    const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
+    const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx;
     const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
 
     // Process Q8_K blocks one by one
@@ -895,11 +1435,11 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         // Pointers to LHS blocks of block_q8_K format
         const block_q8_K * a_ptr = a_ptr_start + (y * nb);
 
-        // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
-        for(int64_t x = 0; x < nc / 8; x++) {
+        // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < nc / 8; x++) {
 
             // Pointers to RHS blocks
-            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
 
             // Master FP accumulators
             __m256 acc_row = _mm256_setzero_ps();
@@ -907,10 +1447,10 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
             for (int64_t b = 0; b < nb; b++) {
 
-                // Load and convert to FP32 delta from block_q8_K
+                // Load and convert to FP32 scale from block_q8_K
                 const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
 
-                // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
+                // Load the scale values for the 8 blocks interleaved in block_q4_Kx8
                 // col_scale_f32 rearranged so as to multiply with appropriate quants
                 const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
                 const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
@@ -918,11 +1458,15 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 __m256i iacc_b = _mm256_setzero_si256();
                 __m256i iacc_min_b = _mm256_setzero_si256();
 
-                // Processes eight sub blocks from each Q2_K in each iteration
-                for(int sb = 0; sb < QK_K / 128; sb++) {
+                const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums));
+                __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1)));
+                q8s = _mm256_permute2f128_si256(q8s, q8s, 0);
 
-                    // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
+                // Processes two sub blocks from each Q4_K in each iteration
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+
+                    // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
                     const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
                     const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
                     const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
@@ -931,982 +1475,482 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
                     const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
 
-                    // 2-bit -> 8-bit
-                    // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
-                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m3b); //B00(0-7) B01(0-7) B02(0-7) B03(0-7)
-                    const __m256i rhs_vec_0123_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 2), m3b); //B20(0-7) B21(0-7) B22(0-7) B23(0-7)
-                    const __m256i rhs_vec_0123_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m3b); //B40(0-7) B41(0-7) B42(0-7) B43(0-7)
-                    const __m256i rhs_vec_0123_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 6), m3b); //B60(0-7) B61(0-7) B62(0-7) B63(0-7)
-
-                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m3b); //B04(0-7) B05(0-7) B06(0-7) B07(0-7)
-                    const __m256i rhs_vec_4567_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 2), m3b); //B24(0-7) B25(0-7) B26(0-7) B27(0-7)
-                    const __m256i rhs_vec_4567_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m3b); //B44(0-7) B45(0-7) B46(0-7) B47(0-7)
-                    const __m256i rhs_vec_4567_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 6), m3b); //B64(0-7) B65(0-7) B66(0-7) B67(0-7)
-
-                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m3b); //B00(8-15) B01(8-15) B02(8-15) B03(8-15)
-                    const __m256i rhs_vec_0123_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 2), m3b); //B20(8-15) B21(8-15) B22(8-15) B23(8-15)
-                    const __m256i rhs_vec_0123_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m3b); //B40(8-15) B41(8-15) B42(8-15) B43(8-15)
-                    const __m256i rhs_vec_0123_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 6), m3b); //B60(8-15) B61(8-15) B62(8-15) B63(8-15)
-
-                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m3b); //B04(8-15) B05(8-15) B06(8-15) B07(8-15)
-                    const __m256i rhs_vec_4567_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 2), m3b); //B24(8-15) B25(8-15) B26(8-15) B27(8-15)
-                    const __m256i rhs_vec_4567_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m3b); //B44(8-15) B45(8-15) B46(8-15) B47(8-15)
-                    const __m256i rhs_vec_4567_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 6), m3b); //B64(8-15) B65(8-15) B66(8-15) B67(8-15)
-
-                    // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
-                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(rhs_raw_vec_0123_2, m3b); //B10(0-7) B11(0-7) B12(0-7) B13(0-7)
-                    const __m256i rhs_vec_0123_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 2), m3b); //B30(0-7) B31(0-7) B32(0-7) B33(0-7)
-                    const __m256i rhs_vec_0123_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m3b); //B50(0-7) B51(0-7) B52(0-7) B53(0-7)
-                    const __m256i rhs_vec_0123_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 6), m3b); //B70(0-7) B71(0-7) B72(0-7) B73(0-7)
-
-                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(rhs_raw_vec_4567_2, m3b); //B14(0-7) B15(0-7) B16(0-7) B17(0-7)
-                    const __m256i rhs_vec_4567_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 2), m3b); //B34(0-7) B35(0-7) B36(0-7) B37(0-7)
-                    const __m256i rhs_vec_4567_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m3b); //B54(0-7) B55(0-7) B56(0-7) B57(0-7)
-                    const __m256i rhs_vec_4567_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 6), m3b); //B74(0-7) B75(0-7) B76(0-7) B77(0-7)
-
-                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(rhs_raw_vec_0123_3, m3b); //B10(8-15) B11(8-15) B12(8-15) B13(8-15)
-                    const __m256i rhs_vec_0123_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 2), m3b); //B30(8-15) B31(8-15) B32(8-15) B33(8-15)
-                    const __m256i rhs_vec_0123_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m3b); //B50(8-15) B51(8-15) B52(8-15) B53(8-15)
-                    const __m256i rhs_vec_0123_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 6), m3b); //B70(8-15) B71(8-15) B72(8-15) B73(8-15)
-
-                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(rhs_raw_vec_4567_3, m3b); //B14(8-15) B15(8-15) B16(8-15) B17(8-15)
-                    const __m256i rhs_vec_4567_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 2), m3b); //B34(8-15) B35(8-15) B36(8-15) B37(8-15)
-                    const __m256i rhs_vec_4567_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m3b); //B54(8-15) B55(8-15) B56(8-15) B57(8-15)
-                    const __m256i rhs_vec_4567_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 6), m3b); //B74(8-15) B75(8-15) B76(8-15) B77(8-15)
-
-                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
-                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
-
-                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
-                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
-                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
-                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
-
-                    // Extract scales which is lower half from mins_and_scales
-                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
-                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
-                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
-                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
-
-                    // Extract mins which is upper half from mins_and_scales
-                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
-                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
-                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
-                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
-
-                    // Scales of sub blocks in the sb loop
-                    // Scales of the 0th sub block from each super block
-                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
-                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
-
-                    // Scales of the 1st sub block from each super block
-                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
-                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
-
-                    // Scales of the 2nd sub block from each super block
-                    __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
-                    __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
-
-                    // Scales of the 3rd sub block from each super block
-                    __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
-                    __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
-
-                    // Scales of the 4th sub block from each super block
-                    __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
-                    __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
-
-                    // Scales of the 5th sub block from each super block
-                    __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
-                    __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
-
-                    // Scales of the 6th sub block from each super block
-                    __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
-                    __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
-
-                    // Scales of the 7th sub block from each super block
-                    __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
-                    __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
-
-                    // Load the sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
-                    __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
-                    __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
-                    __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
-                    __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
-                    __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
-                    __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
-                    __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
-                    __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
-
-                    lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
-                    lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
-                    lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
-                    lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
-                    lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
-                    lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
-                    lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
-                    lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
-
-                    __m256i iacc_0 = _mm256_setzero_si256();
-                    __m256i iacc_1 = _mm256_setzero_si256();
-                    __m256i iacc_2 = _mm256_setzero_si256();
-                    __m256i iacc_3 = _mm256_setzero_si256();
-                    __m256i iacc_4 = _mm256_setzero_si256();
-                    __m256i iacc_5 = _mm256_setzero_si256();
-                    __m256i iacc_6 = _mm256_setzero_si256();
-                    __m256i iacc_7 = _mm256_setzero_si256();
-
-                    // Dot product done within 32 bit lanes and accumulated in the same vector
-                    // First done for 0th sub block and then for seven (1st - 7th) other sub blocks processed for each sb (sb < QK_K/128 loop)                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
-                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
-                    // B0(8-11) B4(8-11) B1(8-11) B5(8-11) B2(8-11) B6(8-11) B3(8-11) B7(8-11) with A0(8-11)
-                    // B0(12-15) B4(12-15) B1(12-15) B5(12-15) B2(12-15) B6(12-15) B3(12-15) B7(12-15) with A0(12-15)
-
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
-
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
-                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
-
-                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
-
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
-
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
-                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
-
-                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
-
-                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_20 ,_mm256_shuffle_epi32(rhs_vec_4567_20, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 0)));
-                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_20, 177) ,rhs_vec_4567_20, 170), _mm256_shuffle_epi32(lhs_vec_2, 85)));
-
-                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_21 ,_mm256_shuffle_epi32(rhs_vec_4567_21, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 170)));
-                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_21, 177) ,rhs_vec_4567_21, 170), _mm256_shuffle_epi32(lhs_vec_2, 255)));
-
-                    iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
-
-                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_30 ,_mm256_shuffle_epi32(rhs_vec_4567_30, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 0)));
-                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_30, 177) ,rhs_vec_4567_30, 170), _mm256_shuffle_epi32(lhs_vec_3, 85)));
-
-                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_31 ,_mm256_shuffle_epi32(rhs_vec_4567_31, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 170)));
-                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_31, 177) ,rhs_vec_4567_31, 170), _mm256_shuffle_epi32(lhs_vec_3, 255)));
-
-                    iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
-
-                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_40 ,_mm256_shuffle_epi32(rhs_vec_4567_40, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 0)));
-                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_40, 177) ,rhs_vec_4567_40, 170), _mm256_shuffle_epi32(lhs_vec_4, 85)));
-
-                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_41 ,_mm256_shuffle_epi32(rhs_vec_4567_41, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 170)));
-                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_41, 177) ,rhs_vec_4567_41, 170), _mm256_shuffle_epi32(lhs_vec_4, 255)));
-
-                    iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
-
-                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_50 ,_mm256_shuffle_epi32(rhs_vec_4567_50, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 0)));
-                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_50, 177) ,rhs_vec_4567_50, 170), _mm256_shuffle_epi32(lhs_vec_5, 85)));
-
-                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_51 ,_mm256_shuffle_epi32(rhs_vec_4567_51, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 170)));
-                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_51, 177) ,rhs_vec_4567_51, 170), _mm256_shuffle_epi32(lhs_vec_5, 255)));
-
-                    iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
-
-                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_60 ,_mm256_shuffle_epi32(rhs_vec_4567_60, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 0)));
-                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_60, 177) ,rhs_vec_4567_60, 170), _mm256_shuffle_epi32(lhs_vec_6, 85)));
-
-                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_61 ,_mm256_shuffle_epi32(rhs_vec_4567_61, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 170)));
-                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_61, 177) ,rhs_vec_4567_61, 170), _mm256_shuffle_epi32(lhs_vec_6, 255)));
-
-                    iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
-
-                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_70 ,_mm256_shuffle_epi32(rhs_vec_4567_70, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 0)));
-                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_70, 177) ,rhs_vec_4567_70, 170), _mm256_shuffle_epi32(lhs_vec_7, 85)));
-
-                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_71 ,_mm256_shuffle_epi32(rhs_vec_4567_71, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 170)));
-                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_71, 177) ,rhs_vec_4567_71, 170), _mm256_shuffle_epi32(lhs_vec_7, 255)));
-
-                    iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
-
-                    // Accumulate the iacc value for one sb
-                    __m256i iacc_sb = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_0, iacc_1), _mm256_add_epi32(iacc_2, iacc_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_4, iacc_5), _mm256_add_epi32(iacc_6, iacc_7)));
-
-                    __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
-                    __m256i q8s = _mm256_castsi128_si256(q8sums);
-                    q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
-
-                    // Broadcast the bsums of the two corresponding subblocks of q8_k
-                    // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
-                    __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
-                    __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
-                    __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
-                    __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
-
-                    __m256i iacc_min_sb = _mm256_add_epi32(_mm256_add_epi32(iacc_min_sb_01, iacc_min_sb_23), _mm256_add_epi32(iacc_min_sb_45,iacc_min_sb_67));
-
-                    // Accumulate for the complete block
-                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
-                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
-                }
-
-                //Multiply-Add with scale values for complete super block
-                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
-                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
-            }
-            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
-            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
-            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
-        }
-    }
-#else
-
-    ggml_gemv_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
-
-#endif
-}
-
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-
-    assert (n % qk == 0);
-    assert (nr % 4 == 0);
-    assert (nc % ncols_interleaved == 0);
-
-    UNUSED(s);
-    UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
-
-#if defined(__AVX2__) || defined(__AVX512F__)
-    {
-        const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
-        const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
-        int64_t b_nb = n / QK4_0;
-        int64_t y = 0;
-        // Mask to mask out nibbles from packed bytes
-        const __m256i m4b = _mm256_set1_epi8(0x0F);
-        const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
-        // Lookup table to convert signed nibbles to signed bytes
-        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
-        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
-        // Permute mask used for easier vector processing at later stages
-        __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
-        int64_t xstart = 0;
-        int anr = nr - nr%16; // Used to align nr with boundary of 16
-    #ifdef __AVX512F__
-        int anc = nc - nc%16; // Used to align nc with boundary of 16
-        // Mask to mask out nibbles from packed bytes expanded to 512 bit length
-        const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
-        // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length
-        __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1);
-
-        // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
-        for (; y < anr / 4; y += 4) {
-
-            const block_q8_0x4 * a_ptrs[4];
-
-            a_ptrs[0] = a_ptr_start + (y * nb);
-            for (int i = 0; i < 3; ++i) {
-                a_ptrs[i + 1] = a_ptrs[i] + nb;
-            }
-
-            // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation
-            for (int64_t x = 0; x < anc / 8; x += 2) {
-
-                const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
-                const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
-
-                // Master FP accumulators
-                __m512 acc_rows[16];
-                for (int i = 0; i < 16; i++) {
-                    acc_rows[i] = _mm512_setzero_ps();
-                }
-
-                for (int64_t b = 0; b < nb; b++) {
-                    // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
-                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
-                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
-                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
-                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
-
-                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
-                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
-                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
-                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
-
-                    // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values
-                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
-                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
-
-                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
-                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
-
-                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
-                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
-                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
-                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
-
-                    // 4-bit -> 8-bit - Sign is maintained
-                    const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
-                    const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
-
-                    const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
-                    const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
-
-                    const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
-                    const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
-
-                    const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
-                    const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
-
-                    // Shuffle pattern one - right side input
-                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
-                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
-
-                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
-                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
-
-                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
-                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
-
-                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
-                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
-
-                    // Shuffle pattern two - right side input
-
-                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
-                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
-
-                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
-                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
-
-                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
-                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
-
-                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
-                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
-
-                    // Scale values - Load the weight scale values of two block_q4_0x8
-                    const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
-
-                    // Process LHS in pairs of rows
-                    for (int rp = 0; rp < 4; rp++) {
-
-                        // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
-                        // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
-                        __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
-                        __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
-                        __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
-                        __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
-                        __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
-                        __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
-                        __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
-                        __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
-                        __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
-                        __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
-                        __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
-                        __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
-
-                        __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
-                        __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
-                        __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
-                        __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
-                        __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
-                        __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
-                        __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
-                        __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
-
-                        // Shuffle pattern one - left side input
-
-                        const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                        const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
-
-                        const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                        const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
-
-                        const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                        const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
-
-                        const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                        const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
-
-                        // Shuffle pattern two - left side input
-
-                        const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                        const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
-
-                        const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                        const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
-
-                        const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                        const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
-
-                        const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                        const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
-
-                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
-                        // Resembles MMLAs into 2x2 matrices in ARM Version
-                        const __m512i zero = _mm512_setzero_epi32();
-                        __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
-                        __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
-                        __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
-                        __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
-                        __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
-                        __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
-                        __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
-                        __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
-
-                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
-                        __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
-                        __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
-                        __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
-                        __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
-
-
-                        // Straighten out to make 4 row vectors
-                        __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
-                        __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
-                        __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
-                        __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
-
-                        // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
-                        const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
-                        const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
-
-                        // Multiply with appropiate scales and accumulate
-                        acc_rows[rp * 4]     = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[rp * 4]);
-                        acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[rp * 4 + 1]);
-                        acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
-                        acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
-                    }
-                }
-
-                // Store the accumulated values
-                for (int i = 0; i < 16; i++) {
-                    _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
-                }
-            }
-        }
-        // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
-        for (; y < nr / 4; y ++) {
-
-            const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
-
-            // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation
-            for (int64_t x = 0; x < anc / 8; x += 2) {
-
-                const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
-                const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
-
-                // Master FP accumulators
-                __m512 acc_rows[4];
-                for (int i = 0; i < 4; i++) {
-                    acc_rows[i] = _mm512_setzero_ps();
-                }
-
-                for (int64_t b = 0; b < nb; b++) {
-                    // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
-                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
-                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
-                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
-                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
-
-                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
-                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
-                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
-                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
-
-                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
-                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
-                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
-
-                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
-                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+                    // 4-bit -> 8-bit
+                    // Values of the first sub block of eight block_q4_K structures for the sb loop
+                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b);
+                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b);
+                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b);
+                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b);
+                    const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b);
+                    const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b);
+                    const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b);
+                    const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b);
 
-                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
-                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
-                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
-                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+                    // Values of the second sub block of eight block_q4_K structures when sb = 1
+                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b);
+                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b);
+                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b);
+                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b);
+                    const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b);
+                    const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b);
+                    const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b);
+                    const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b);
 
-                    // 4-bit -> 8-bit - Sign is maintained
-                    const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
-                    const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
+                    uint32_t utmp_0[4], utmp_1[4];
 
-                    const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
-                    const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
+                    // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together
+                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
+                    utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp_0[1] & kmask1;
+                    utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
+                    utmp_0[2] = uaux_0;
+                    utmp_0[0] &= kmask1;
 
-                    const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
-                    const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
+                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+                    memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
+                    utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_1 = utmp_1[1] & kmask1;
+                    utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
+                    utmp_1[2] = uaux_1;
+                    utmp_1[0] &= kmask1;
 
-                    const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
-                    const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
+                    // Scales of first sub block in the sb loop
+                    const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
+                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask);
+                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
 
-                    // Shuffle pattern one - right side input
-                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
-                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+                    // Scales of second sub block in the sb loop
+                    __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
+                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask);
+                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
 
-                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
-                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+                    // Mins of first and second sub block of Q4_K block are arranged side by side
+                    __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
 
-                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
-                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+                    // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
+                    __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64)));
+                    __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64)));
+                    __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64)));
+                    __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64)));
 
-                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
-                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+                    lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0);
+                    lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0);
+                    lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0);
+                    lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
 
-                    // Shuffle pattern two - right side input
+                    // Dot product done within 32 bit lanes and accumulated in the same vector
+                    // First done for first sub block and thenn for second sub block in each sb
+                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+                    // ...........................................................................
+                    // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
 
-                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
-                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
 
-                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
-                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+                    __m256i iacc_0 = _mm256_setzero_si256();
+                    __m256i iacc_1 = _mm256_setzero_si256();
 
-                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
-                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85)));
 
-                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
-                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255)));
 
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85)));
 
-                    // Scale values - Load the weight scale values of two block_q4_0x8
-                    const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255)));
 
-                    // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
-                    // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
-                    __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
-                    __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
-                    __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
-                    __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
-                    __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
-                    __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
-                    __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
-                    __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
-                    __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
-                    __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
-                    __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
-                    __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
 
-                    __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
-                    __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
-                    __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
-                    __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
-                    __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
-                    __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
-                    __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
-                    __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85)));
 
-                    // Shuffle pattern one - left side input
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255)));
 
-                    const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                    const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85)));
 
-                    const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                    const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255)));
 
-                    const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                    const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
 
-                    const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                    const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+                    // Accumulate the iacc value for one sb
+                    __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1);
 
-                    // Shuffle pattern two - left side input
+                    // Broadcast the bsums of the two sub blocks  of the iteration of Q8_K across the vector
+                    // Multiply-Add with corresponding mins of Q4_Kx8 with bsums
+                    __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0);
+                    __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01);
+                    q8s = _mm256_bsrli_epi128(q8s, 4);
 
-                    const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                    const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+                    // Accumulate for the complete block
+                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
+                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+                }
 
-                    const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                    const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+                // Multiply-Add with scale values for the complete super block
+                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
 
-                    const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                    const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+            }
 
-                    const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                    const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+        }
+    }
 
-                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
-                    // Resembles MMLAs into 2x2 matrices in ARM Version
-                    const __m512i zero = _mm512_setzero_epi32();
-                    __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
-                    __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
-                    __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
-                    __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
-                    __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
-                    __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
-                    __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
-                    __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
+#else
+    UNUSED(kmask1);
+    UNUSED(kmask2);
+    UNUSED(kmask3);
+    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+#endif
+}
 
-                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
-                    __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
-                    __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
-                    __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
-                    __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__)
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
 
+    gemv_q4_b32_8x8_q8_0_lut_avx<block_iq4_nlx8>(n, s, bs, vx, vy, nr, nc, signextendlut);
 
-                    // Straighten out to make 4 row vectors
-                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
-                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
-                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
-                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+    return;
+#endif
 
-                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
-                    const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
-                    const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+    ggml_gemv_iq4_nl_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
 
-                    // Multiply with appropiate scales and accumulate
-                    acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[0]);
-                    acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[1]);
-                    acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
-                    acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
-                }
+void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
 
-                // Store the accumulated values
-                for (int i = 0; i < 4; i++) {
-                    _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
-                }
-            }
-        }
-        if (anc != nc) {
-            xstart = anc/8;
-            y = 0;
-        }
-    #endif // __AVX512F__
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
 
-        // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
 
-        for (; y < anr / 4; y += 4) {
-            const block_q8_0x4 * a_ptrs[4];
+#if defined(__AVX2__)
+    // Lookup table to convert signed nibbles to signed bytes
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+    // Shuffle masks to rearrange delta values to multiply with appropriate scales
+    __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+    // Permute mask used for easier vector processing at later stages
+    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
 
-            a_ptrs[0] = a_ptr_start + (y * nb);
-            for (int i = 0; i < 3; ++i) {
-                a_ptrs[i + 1] = a_ptrs[i] + nb;
-            }
+    const __m256i m3b = _mm256_set1_epi8(3);
+    const __m128i m4b_sse = _mm_set1_epi8(0xF);
 
-            // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
-            for (int64_t x = xstart; x < nc / 8; x++) {
+    //Mask to get appropriate scales
+    __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
+    __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
 
-                const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+    int64_t b_nb = n / QK_K;
 
-                // Master FP accumulators
-                __m256 acc_rows[16];
-                for (int i = 0; i < 16; i++) {
-                    acc_rows[i] = _mm256_setzero_ps();
-                }
+    const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
+    const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
 
-                for (int64_t b = 0; b < nb; b++) {
-                    // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
-                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
-                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
-                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+    // Process Q8_K blocks one by one
+    for (int64_t y = 0; y < nr; y++) {
 
-                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
-                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
-                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+        // Pointers to LHS blocks of block_q8_K format
+        const block_q8_K * a_ptr = a_ptr_start + (y * nb);
 
-                    // 4-bit -> 8-bit - Sign is maintained
-                    const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
-                    const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+        // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
+        for(int64_t x = 0; x < nc / 8; x++) {
 
-                    const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
-                    const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+            // Pointers to RHS blocks
+            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
 
-                    const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
-                    const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+            // Master FP accumulators
+            __m256 acc_row = _mm256_setzero_ps();
+            __m256 acc_min_rows = _mm256_setzero_ps();
 
-                    const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
-                    const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+            for (int64_t b = 0; b < nb; b++) {
 
-                    // Shuffle pattern one - right side input
-                    const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
-                    const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+                // Load and convert to FP32 delta from block_q8_K
+                const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
 
-                    const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
-                    const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+                // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
+                // col_scale_f32 rearranged so as to multiply with appropriate quants
+                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
+                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
 
-                    const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
-                    const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+                __m256i iacc_b = _mm256_setzero_si256();
+                __m256i iacc_min_b = _mm256_setzero_si256();
 
-                    const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
-                    const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+                // Processes eight sub blocks from each Q2_K in each iteration
+                for(int sb = 0; sb < QK_K / 128; sb++) {
 
-                    // Shuffle pattern two - right side input
+                    // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
+                    const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
+                    const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
+                    const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
 
-                    const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
-                    const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+                    // 2-bit -> 8-bit
+                    // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
+                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m3b); //B00(0-7) B01(0-7) B02(0-7) B03(0-7)
+                    const __m256i rhs_vec_0123_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 2), m3b); //B20(0-7) B21(0-7) B22(0-7) B23(0-7)
+                    const __m256i rhs_vec_0123_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m3b); //B40(0-7) B41(0-7) B42(0-7) B43(0-7)
+                    const __m256i rhs_vec_0123_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 6), m3b); //B60(0-7) B61(0-7) B62(0-7) B63(0-7)
 
-                    const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
-                    const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m3b); //B04(0-7) B05(0-7) B06(0-7) B07(0-7)
+                    const __m256i rhs_vec_4567_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 2), m3b); //B24(0-7) B25(0-7) B26(0-7) B27(0-7)
+                    const __m256i rhs_vec_4567_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m3b); //B44(0-7) B45(0-7) B46(0-7) B47(0-7)
+                    const __m256i rhs_vec_4567_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 6), m3b); //B64(0-7) B65(0-7) B66(0-7) B67(0-7)
 
-                    const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
-                    const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m3b); //B00(8-15) B01(8-15) B02(8-15) B03(8-15)
+                    const __m256i rhs_vec_0123_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 2), m3b); //B20(8-15) B21(8-15) B22(8-15) B23(8-15)
+                    const __m256i rhs_vec_0123_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m3b); //B40(8-15) B41(8-15) B42(8-15) B43(8-15)
+                    const __m256i rhs_vec_0123_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 6), m3b); //B60(8-15) B61(8-15) B62(8-15) B63(8-15)
 
-                    const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
-                    const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m3b); //B04(8-15) B05(8-15) B06(8-15) B07(8-15)
+                    const __m256i rhs_vec_4567_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 2), m3b); //B24(8-15) B25(8-15) B26(8-15) B27(8-15)
+                    const __m256i rhs_vec_4567_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m3b); //B44(8-15) B45(8-15) B46(8-15) B47(8-15)
+                    const __m256i rhs_vec_4567_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 6), m3b); //B64(8-15) B65(8-15) B66(8-15) B67(8-15)
 
-                    // Scale values - Load the wight scale values of block_q4_0x8
-                    const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                    // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
+                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(rhs_raw_vec_0123_2, m3b); //B10(0-7) B11(0-7) B12(0-7) B13(0-7)
+                    const __m256i rhs_vec_0123_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 2), m3b); //B30(0-7) B31(0-7) B32(0-7) B33(0-7)
+                    const __m256i rhs_vec_0123_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m3b); //B50(0-7) B51(0-7) B52(0-7) B53(0-7)
+                    const __m256i rhs_vec_0123_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 6), m3b); //B70(0-7) B71(0-7) B72(0-7) B73(0-7)
 
-                    // Process LHS in groups of four
-                    for (int rp = 0; rp < 4; rp++) {
-                        // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
-                        // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
-                        __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
-                        __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
-                        __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
-                        __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
-                        __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
-                        __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
-                        __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
-                        __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
-                        __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
-                        __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
-                        __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
-                        __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(rhs_raw_vec_4567_2, m3b); //B14(0-7) B15(0-7) B16(0-7) B17(0-7)
+                    const __m256i rhs_vec_4567_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 2), m3b); //B34(0-7) B35(0-7) B36(0-7) B37(0-7)
+                    const __m256i rhs_vec_4567_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m3b); //B54(0-7) B55(0-7) B56(0-7) B57(0-7)
+                    const __m256i rhs_vec_4567_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 6), m3b); //B74(0-7) B75(0-7) B76(0-7) B77(0-7)
 
-                        // Shuffle pattern one - left side input
-                        const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                        const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(rhs_raw_vec_0123_3, m3b); //B10(8-15) B11(8-15) B12(8-15) B13(8-15)
+                    const __m256i rhs_vec_0123_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 2), m3b); //B30(8-15) B31(8-15) B32(8-15) B33(8-15)
+                    const __m256i rhs_vec_0123_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m3b); //B50(8-15) B51(8-15) B52(8-15) B53(8-15)
+                    const __m256i rhs_vec_0123_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 6), m3b); //B70(8-15) B71(8-15) B72(8-15) B73(8-15)
 
-                        const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                        const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(rhs_raw_vec_4567_3, m3b); //B14(8-15) B15(8-15) B16(8-15) B17(8-15)
+                    const __m256i rhs_vec_4567_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 2), m3b); //B34(8-15) B35(8-15) B36(8-15) B37(8-15)
+                    const __m256i rhs_vec_4567_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m3b); //B54(8-15) B55(8-15) B56(8-15) B57(8-15)
+                    const __m256i rhs_vec_4567_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 6), m3b); //B74(8-15) B75(8-15) B76(8-15) B77(8-15)
 
-                        const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                        const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+                    //s00 m00  s01 m01   s10 m10  s11 m11  s20 m20  s21 m21   s30 m30  s31 m31  s40 m40  s41 m41   s50 m50  s51 m51  s60 m60  s61 m61   s70 m70  s71 m71
 
-                        const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                        const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
 
-                        // Shuffle pattern two - left side input
-                        const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                        const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+                    // Extract scales which is lower half from mins_and_scales
+                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
 
-                        const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                        const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+                    // Extract mins which is upper half from mins_and_scales
+                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
 
-                        const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                        const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+                    // Scales of sub blocks in the sb loop
+                    // Scales of the 0th sub block from each super block
+                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
+                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
 
-                        const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                        const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+                    // Scales of the 1st sub block from each super block
+                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
+                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
 
-                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
-                        // Resembles MMLAs into 2x2 matrices in ARM Version
-                        const __m256i zero = _mm256_setzero_si256();
-                        __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
-                        __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
-                        __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
-                        __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
-                        __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
-                        __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
-                        __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
-                        __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+                    // Scales of the 2nd sub block from each super block
+                    __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
+                    __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
 
-                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
-                        __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
-                        __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
-                        __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
-                        __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+                    // Scales of the 3rd sub block from each super block
+                    __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
+                    __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
 
-                        // Straighten out to make 4 row vectors
-                        __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
-                        __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
-                        __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
-                        __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+                    // Scales of the 4th sub block from each super block
+                    __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
+                    __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
 
-                        // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
-                        const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
+                    // Scales of the 5th sub block from each super block
+                    __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
+                    __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
 
-                        // Multiply with appropiate scales and accumulate
-                        acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
-                        acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
-                        acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
-                        acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32,  255)), acc_rows[rp * 4 + 3]);
-                    }
-                }
+                    // Scales of the 6th sub block from each super block
+                    __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
+                    __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
 
-                // Store the accumulated values
-                for (int i = 0; i < 16; i++) {
-                    _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
-                }
-            }
-        }
+                    // Scales of the 7th sub block from each super block
+                    __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
+                    __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
 
-        // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
-        for (; y < nr / 4; y ++) {
+                    // Load the sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
+                    __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
+                    __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
+                    __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
+                    __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
+                    __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
+                    __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
+                    __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
+                    __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
 
-            const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+                    lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
+                    lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
+                    lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
+                    lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
+                    lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
+                    lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
+                    lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
+                    lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
 
-            // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-            for (int64_t x = xstart; x < nc / 8; x++) {
+                    __m256i iacc_0 = _mm256_setzero_si256();
+                    __m256i iacc_1 = _mm256_setzero_si256();
+                    __m256i iacc_2 = _mm256_setzero_si256();
+                    __m256i iacc_3 = _mm256_setzero_si256();
+                    __m256i iacc_4 = _mm256_setzero_si256();
+                    __m256i iacc_5 = _mm256_setzero_si256();
+                    __m256i iacc_6 = _mm256_setzero_si256();
+                    __m256i iacc_7 = _mm256_setzero_si256();
 
-                const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+                    // Dot product done within 32 bit lanes and accumulated in the same vector
+                    // First done for 0th sub block and then for seven (1st - 7th) other sub blocks processed for each sb (sb < QK_K/128 loop)                    // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+                    // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+                    // B0(8-11) B4(8-11) B1(8-11) B5(8-11) B2(8-11) B6(8-11) B3(8-11) B7(8-11) with A0(8-11)
+                    // B0(12-15) B4(12-15) B1(12-15) B5(12-15) B2(12-15) B6(12-15) B3(12-15) B7(12-15) with A0(12-15)
 
-                // Master FP accumulators
-                __m256 acc_rows[4];
-                for (int i = 0; i < 4; i++) {
-                    acc_rows[i] = _mm256_setzero_ps();
-                }
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
 
-                for (int64_t b = 0; b < nb; b++) {
-                    // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
-                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
-                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
-                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
-                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
+                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
 
-                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
-                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
-                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
-                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
 
-                    // 4-bit -> 8-bit - Sign is maintained
-                    const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b));  //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
-                    const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b));  //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
 
-                    const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b));  //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
-                    const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b));  //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
+                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
 
-                    const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b));  //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
-                    const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b));  //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
 
-                    const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b));  //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
-                    const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b));  //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_20 ,_mm256_shuffle_epi32(rhs_vec_4567_20, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 0)));
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_20, 177) ,rhs_vec_4567_20, 170), _mm256_shuffle_epi32(lhs_vec_2, 85)));
 
-                    // Shuffle pattern one - right side input
-                    const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
-                    const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_21 ,_mm256_shuffle_epi32(rhs_vec_4567_21, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 170)));
+                    iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_21, 177) ,rhs_vec_4567_21, 170), _mm256_shuffle_epi32(lhs_vec_2, 255)));
 
-                    const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
-                    const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+                    iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
 
-                    const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
-                    const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_30 ,_mm256_shuffle_epi32(rhs_vec_4567_30, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 0)));
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_30, 177) ,rhs_vec_4567_30, 170), _mm256_shuffle_epi32(lhs_vec_3, 85)));
 
-                    const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
-                    const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_31 ,_mm256_shuffle_epi32(rhs_vec_4567_31, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 170)));
+                    iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_31, 177) ,rhs_vec_4567_31, 170), _mm256_shuffle_epi32(lhs_vec_3, 255)));
 
-                    // Shuffle pattern two - right side input
+                    iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
 
-                    const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
-                    const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_40 ,_mm256_shuffle_epi32(rhs_vec_4567_40, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 0)));
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_40, 177) ,rhs_vec_4567_40, 170), _mm256_shuffle_epi32(lhs_vec_4, 85)));
 
-                    const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
-                    const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_41 ,_mm256_shuffle_epi32(rhs_vec_4567_41, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 170)));
+                    iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_41, 177) ,rhs_vec_4567_41, 170), _mm256_shuffle_epi32(lhs_vec_4, 255)));
 
-                    const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
-                    const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+                    iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
 
-                    const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
-                    const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_50 ,_mm256_shuffle_epi32(rhs_vec_4567_50, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 0)));
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_50, 177) ,rhs_vec_4567_50, 170), _mm256_shuffle_epi32(lhs_vec_5, 85)));
 
-                    // Scale values - Load the wight scale values of block_q4_0x8
-                    const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_51 ,_mm256_shuffle_epi32(rhs_vec_4567_51, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 170)));
+                    iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_51, 177) ,rhs_vec_4567_51, 170), _mm256_shuffle_epi32(lhs_vec_5, 255)));
 
-                    // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
-                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
-                    __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
-                    __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
-                    __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
-                    __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
-                    __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
-                    __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
-                    __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
-                    __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
-                    __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
-                    __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
-                    __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
-                    __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+                    iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
 
-                    // Shuffle pattern one - left side input
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_60 ,_mm256_shuffle_epi32(rhs_vec_4567_60, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 0)));
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_60, 177) ,rhs_vec_4567_60, 170), _mm256_shuffle_epi32(lhs_vec_6, 85)));
 
-                    const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                    const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_61 ,_mm256_shuffle_epi32(rhs_vec_4567_61, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 170)));
+                    iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_61, 177) ,rhs_vec_4567_61, 170), _mm256_shuffle_epi32(lhs_vec_6, 255)));
 
-                    const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                    const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+                    iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
 
-                    const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                    const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_70 ,_mm256_shuffle_epi32(rhs_vec_4567_70, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 0)));
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_70, 177) ,rhs_vec_4567_70, 170), _mm256_shuffle_epi32(lhs_vec_7, 85)));
 
-                    const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                    const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_71 ,_mm256_shuffle_epi32(rhs_vec_4567_71, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 170)));
+                    iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_71, 177) ,rhs_vec_4567_71, 170), _mm256_shuffle_epi32(lhs_vec_7, 255)));
 
-                    // Shuffle pattern two - left side input
+                    iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
 
-                    const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                    const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+                    // Accumulate the iacc value for one sb
+                    __m256i iacc_sb = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_0, iacc_1), _mm256_add_epi32(iacc_2, iacc_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_4, iacc_5), _mm256_add_epi32(iacc_6, iacc_7)));
 
-                    const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                    const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+                    __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
+                    __m256i q8s = _mm256_castsi128_si256(q8sums);
+                    q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
 
-                    const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                    const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+                    // Broadcast the bsums of the two corresponding subblocks of q8_k
+                    // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
+                    __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
+                    __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
+                    __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
+                    __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
 
-                    const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                    const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+                    __m256i iacc_min_sb = _mm256_add_epi32(_mm256_add_epi32(iacc_min_sb_01, iacc_min_sb_23), _mm256_add_epi32(iacc_min_sb_45,iacc_min_sb_67));
 
-                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
-                    // Resembles MMLAs into 2x2 matrices in ARM Version
-                    const __m256i zero = _mm256_setzero_si256();
-                    __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
-                    __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
-                    __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
-                    __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
-                    __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
-                    __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
-                    __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
-                    __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+                    // Accumulate for the complete block
+                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
+                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+                }
 
-                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
-                    __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
-                    __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
-                    __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
-                    __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+                //Multiply-Add with scale values for complete super block
+                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
+            }
+            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+        }
+    }
+#else
 
+    ggml_gemv_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 
-                    // Straighten out to make 4 row vectors
-                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
-                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
-                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
-                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+#endif
+}
 
-                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
-                    const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+    {
+        // Lookup table to convert signed nibbles to signed bytes
+        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
 
-                    // Multiply with appropiate scales and accumulate
-                    acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
-                    acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
-                    acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
-                    acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
-                }
+        gemm_q4_b32_8x8_q8_0_lut_avx<block_q4_0x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
 
-                // Store the accumulated values
-                for (int i = 0; i < 4; i++) {
-                    _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
-                }
-            }
-        }
         return;
     }
+#endif // defined(__AVX2__) || defined(__AVX512F__)
 
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
     ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
 }
 
@@ -3364,6 +3408,21 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 #endif
 }
 
+void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+    {
+        __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl));
+        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+
+        gemm_q4_b32_8x8_q8_0_lut_avx<block_iq4_nlx8>(n, s, bs, vx, vy, nr, nc, signextendlut);
+
+        return;
+    }
+#endif // defined(__AVX2__) || defined(__AVX512F__)
+
+    ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK_K;
     const int nb = n / qk;
diff --git a/ggml/src/ggml-cpu/ggml-cpu-impl.h b/ggml/src/ggml-cpu/ggml-cpu-impl.h
index d839cf5c55e81..1f6844e16cd34 100644
--- a/ggml/src/ggml-cpu/ggml-cpu-impl.h
+++ b/ggml/src/ggml-cpu/ggml-cpu-impl.h
@@ -486,6 +486,14 @@ inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
     return v_abo + v_abe;
 }
 
+/**
+ * @see https://github.com/ggml-org/llama.cpp/pull/14037
+ */
+inline float vec_hsum(float32x4_t v) {
+    float32x4_t v_temp = v + vec_reve(v);
+    return v_temp[0] + v_temp[1];
+}
+
 inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
     const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
     return acc + (vec_unpackh(p) + vec_unpackl(p));
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index d89cd8f4ef652..0d5d3a3440aaf 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -1880,6 +1880,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_conv_2d(params, tensor);
             } break;
+        case GGML_OP_CONV_3D:
+            {
+                ggml_compute_forward_conv_3d(params, tensor);
+            } break;
         case GGML_OP_CONV_2D_DW:
             {
                 ggml_compute_forward_conv_2d_dw(params, tensor);
@@ -2022,6 +2026,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
                 ggml_compute_forward_opt_step_adamw(params, tensor);
             }
             break;
+        case GGML_OP_OPT_STEP_SGD:
+            {
+                ggml_compute_forward_opt_step_sgd(params, tensor);
+            }
+            break;
         case GGML_OP_NONE:
             {
                 // nop
@@ -2247,6 +2256,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
         case GGML_OP_CONV_2D:
+        case GGML_OP_CONV_3D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
@@ -2325,6 +2335,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             {
                 n_tasks = n_threads;
             } break;
@@ -2767,6 +2778,7 @@ struct ggml_cplan ggml_graph_plan(
                         }
                     } break;
                 case GGML_OP_CONV_2D:
+                case GGML_OP_CONV_3D:
                     {
                         cur = GGML_IM2COL_WORK_SIZE;
                     } break;
diff --git a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
index 3a513a55d7654..dff8fa244a1c9 100644
--- a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
+++ b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
@@ -259,7 +259,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                 const int64_t m_start      = 0;
 
                 const int64_t n_step      = static_cast<int64_t>(kernel->get_n_step());
-                const int64_t num_threads = KAI_MIN(n / n_step, nth);
+                int64_t num_threads       = KAI_MIN(n / n_step, nth);
+                if (num_threads <= 0) {
+                    num_threads = 1;
+                }
 
                 if (ith < num_threads) {
                     const int64_t num_n_per_thread0   = round_down(n / num_threads, n_step);
@@ -309,7 +312,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         GGML_ASSERT(kernel);
 
         const int ith = params->ith;
-        const int nth = params->nth;
+        const int nth_raw = params->nth;
+        const int nth = nth_raw > 0 ? nth_raw : 1;
 
         const size_t k = ne00;
         const size_t m = ne11;
@@ -327,9 +331,12 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
         const size_t n_start = ith * num_n_per_thread;
 
-        size_t n_to_process = num_n_per_thread;
-        if ((n_start + n_to_process) > n) {
-            n_to_process = n - n_start;
+        size_t n_to_process = 0;
+        if (n_start < n) {
+            n_to_process = num_n_per_thread;
+            if ((n_start + n_to_process) > n) {
+                n_to_process = n - n_start;
+            }
         }
 
         // Calculate number of columns to be processed per thread
@@ -361,8 +368,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         const void* lhs_ptr            = (const void*)((const char *)lhs_packed + lhs_packed_offset);
         float *dst_ptr                 = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
 
-        variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
-                           sizeof(float), -FLT_MAX, FLT_MAX);
+        if (n_to_process > 0) {
+            variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+                               sizeof(float), -FLT_MAX, FLT_MAX);
+        }
 
         return true;
     }
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 854f1c2b49647..460367cca09e9 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -7207,6 +7207,148 @@ void ggml_compute_forward_conv_2d(
     ggml_compute_forward_conv_2d_impl(params, src0, src1, dst, src0->type);
 }
 
+// ggml_compute_forward_conv_3d
+
+static void ggml_compute_forward_conv_3d_impl(const ggml_compute_params * params,
+                                              const ggml_tensor *         kernel,
+                                              const ggml_tensor *         src,
+                                              ggml_tensor *               dst,
+                                              ggml_type                   kernel_type) {
+
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(kernel_type == GGML_TYPE_F16 || kernel_type == GGML_TYPE_F32);
+    GGML_ASSERT(kernel->type == kernel_type);
+
+    const ggml_type_traits * traits = ggml_get_type_traits(kernel_type);
+
+    const int32_t s0 = dst->op_params[0];
+    const int32_t s1 = dst->op_params[1];
+    const int32_t s2 = dst->op_params[2];
+    const int32_t p0 = dst->op_params[3];
+    const int32_t p1 = dst->op_params[4];
+    const int32_t p2 = dst->op_params[5];
+    const int32_t d0 = dst->op_params[6];
+    const int32_t d1 = dst->op_params[7];
+    const int32_t d2 = dst->op_params[8];
+    const int32_t c  = dst->op_params[9];
+    const int32_t n  = dst->op_params[10];
+    const int32_t oc = dst->op_params[11];
+
+    const int64_t src_w = src->ne[0];
+    const int64_t src_h = src->ne[1];
+    const int64_t src_d = src->ne[2];
+    const int64_t knl_w = kernel->ne[0];
+    const int64_t knl_h = kernel->ne[1];
+    const int64_t knl_d = kernel->ne[2];
+    const int64_t dst_w = dst->ne[0];
+    const int64_t dst_h = dst->ne[1];
+    const int64_t dst_d = dst->ne[2];
+
+    const float * src_data = (float *) src->data;
+    void  * knl_data       = kernel->data;
+    float * dst_data       = (float *) dst->data;
+
+    const int64_t knl_n_per_channel = knl_w * knl_h * knl_d;
+    const int64_t knl_n_total       = knl_n_per_channel * c;
+    const int64_t patch_total       = n * dst_w * dst_h * dst_d;
+
+    const int64_t space_per_patch   = knl_n_total * traits->type_size + oc * sizeof(float);
+    const int64_t batch_size        = params->wsize / space_per_patch;
+    const int64_t patches_per_batch = batch_size > 8 ? (batch_size / 8) * 8 : batch_size;
+    const int64_t batch_n           = (patch_total + patches_per_batch - 1) / patches_per_batch;
+
+    GGML_ASSERT(patches_per_batch > 0 && batch_size >= 1);
+
+    void * tmp = params->wdata;
+
+    for (int64_t batch_i = 0; batch_i < batch_n; ++batch_i) {
+        const int64_t patch_start_batch = batch_i * patches_per_batch;
+        const int64_t patch_end_batch   = std::min(patch_start_batch + patches_per_batch, patch_total);
+        const int64_t patch_n_in_batch  = patch_end_batch - patch_start_batch;
+
+        const int64_t patch_per_thread  = (patch_n_in_batch + params->nth - 1) / params->nth;
+        const int64_t patch_start       = patch_start_batch + params->ith * patch_per_thread;
+        const int64_t patch_end         = std::min(patch_start + patch_per_thread, patch_end_batch);
+
+        for (int64_t p = patch_start; p < patch_end; ++p) {
+            const int64_t p_in_batch = p % (dst_w * dst_h * dst_d);
+            const int64_t p_in_depth = p_in_batch % (dst_w * dst_h);
+            const int64_t batch_idx  = p / (dst_w * dst_h * dst_d);
+            const int64_t dst_z      = p_in_batch / (dst_w * dst_h);
+            const int64_t dst_y      = p_in_depth / dst_w;
+            const int64_t dst_x      = p_in_depth % dst_w;
+
+            char * dst_row = (char *) tmp + (p % patches_per_batch) * knl_n_total * traits->type_size;
+
+            for (int64_t ic = 0; ic < c; ++ic) {
+                for (int64_t kz = 0; kz < knl_d; ++kz) {
+                    for (int64_t ky = 0; ky < knl_h; ++ky) {
+                        for (int64_t kx = 0; kx < knl_w; ++kx) {
+                            const int64_t sz = dst_z * s2 + kz * d2 - p2;
+                            const int64_t sy = dst_y * s1 + ky * d1 - p1;
+                            const int64_t sx = dst_x * s0 + kx * d0 - p0;
+
+                            int64_t dst_idx = ic * knl_n_per_channel + kz * (knl_h * knl_w) + ky * knl_w + kx;
+
+                            float src_val;
+                            if (sz < 0 || sz >= src_d || sy < 0 || sy >= src_h || sx < 0 || sx >= src_w) {
+                                src_val = 0.0f;
+                            } else {
+                                const int64_t cn_idx = batch_idx * c + ic;
+                                const float * src_ptr = (const float *)((const char *)src_data + sx*src->nb[0] + sy*src->nb[1] + sz*src->nb[2] + cn_idx*src->nb[3]);
+                                src_val = *src_ptr;
+                            }
+
+                            char * element_ptr = dst_row + dst_idx * traits->type_size;
+                            if (kernel_type == GGML_TYPE_F32) {
+                                *(float *)element_ptr = src_val;
+                            } else if (kernel_type == GGML_TYPE_F16) {
+                                *(ggml_fp16_t *)element_ptr = GGML_CPU_FP32_TO_FP16(src_val);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        ggml_barrier(params->threadpool);
+
+        float * gemm_output = (float *) ((char *) tmp + patches_per_batch * knl_n_total * traits->type_size);
+        ggml_call_mul_mat(kernel_type, params, patch_n_in_batch, oc, knl_n_total, tmp, knl_data, gemm_output);
+
+        ggml_barrier(params->threadpool);
+
+        const int64_t permute_per_thread = (patch_n_in_batch + params->nth - 1) / params->nth;
+        const int64_t permute_start = params->ith * permute_per_thread;
+        const int64_t permute_end = std::min(permute_start + permute_per_thread, patch_n_in_batch);
+
+        for (int64_t i = permute_start; i < permute_end; ++i) {
+            const int64_t p = patch_start_batch + i;
+            const int64_t p_in_batch = p % (dst_w * dst_h * dst_d);
+            const int64_t p_in_depth = p_in_batch % (dst_w * dst_h);
+            const int64_t batch_idx  = p / (dst_w * dst_h * dst_d);
+            const int64_t dst_z      = p_in_batch / (dst_w * dst_h);
+            const int64_t dst_y      = p_in_depth / dst_w;
+            const int64_t dst_x      = p_in_depth % dst_w;
+
+            for (int64_t ioc = 0; ioc < oc; ++ioc) {
+                const float value = gemm_output[i * oc + ioc];
+                const int64_t ocn_idx = batch_idx * oc + ioc;
+                float * dst_ptr = (float *)((char *)dst_data + dst_x*dst->nb[0] + dst_y*dst->nb[1] + dst_z*dst->nb[2] + ocn_idx*dst->nb[3]);
+                *dst_ptr = value;
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_conv_3d(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    ggml_compute_forward_conv_3d_impl(params, src0, src1, dst, src0->type);
+}
+
 // ggml_compute_forward_conv_transpose_2d
 
 void ggml_compute_forward_conv_transpose_2d(
@@ -10330,6 +10472,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const int ir1 = MIN(ir0 + dr, nr);
 
     const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
+
     const float alpha  = adamw_params_ptr[0];
     const float beta1  = adamw_params_ptr[1];
     const float beta2  = adamw_params_ptr[2];
@@ -10337,7 +10480,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const float wd     = adamw_params_ptr[4];
     const float beta1h = adamw_params_ptr[5];
     const float beta2h = adamw_params_ptr[6];
-
+    const float keep   = 1.f - alpha * wd;
     for (int ir = ir0; ir < ir1; ++ir) {
         const int64_t i03 = ir/(ne02*ne01);
         const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
@@ -10360,7 +10503,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
             // The weight decay is applied independently of the Adam momenta m and v.
             // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
             // See: https://arxiv.org/pdf/1711.05101v3.pdf
-            w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
+            w[i00] = w[i00] * keep - alpha * mh / vh;
         }
     }
 }
@@ -10382,3 +10525,63 @@ void ggml_compute_forward_opt_step_adamw(
             }
     }
 }
+
+static void ggml_compute_forward_opt_step_sgd_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0       = dst->src[0];
+    const ggml_tensor * src0_grad  = dst->src[1];
+    const ggml_tensor * sgd_params = dst->src[2];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_nelements(sgd_params) == 2);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr = ggml_nrows(src0);
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // rows per thread
+    const int dr = (nr + nth - 1) / nth;
+
+    // row range for this thread
+    const int ir0 = dr * ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    // using adamw param subset we care about - alpha, wd - could have a separate struct
+    const float * sgd_params_ptr   = ggml_get_data_f32(sgd_params);
+    const float   alpha            = sgd_params_ptr[0];
+    const float   keep             = 1.f - alpha * sgd_params_ptr[1];
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir / (ne02 * ne01);
+        const int64_t i02 = (ir - i03 * ne02 * ne01) / ne01;
+        const int64_t i01 = (ir - i03 * ne02 * ne01 - i02 * ne01);
+
+        const size_t offset = i03 * nb03 + i02 * nb02 + i01 * nb01;
+
+        float *       w = (float *) ((char *) src0->data + offset);                   // weight
+        const float * g = (const float *) ((const char *) src0_grad->data + offset);  // grad
+
+        for (int i00 = 0; i00 < ne00; ++i00) {
+            w[i00] = w[i00] * keep - alpha * g[i00];
+        }
+    }
+}
+
+void ggml_compute_forward_opt_step_sgd(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_opt_step_sgd_f32(params, dst);
+            }
+            break;
+        default:
+            {
+                GGML_ABORT("fatal error - sgd is F32 only");
+            }
+    }
+}
diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h
index f154afb462498..d0ea83843b544 100644
--- a/ggml/src/ggml-cpu/ops.h
+++ b/ggml/src/ggml-cpu/ops.h
@@ -70,6 +70,7 @@ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * p
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_3d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -107,7 +108,7 @@ void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params *
 void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-
+void ggml_compute_forward_opt_step_sgd(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 #ifdef __cplusplus
 }
 #endif
diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp
index 2583aefae47a2..f531d21e23224 100644
--- a/ggml/src/ggml-cpu/repack.cpp
+++ b/ggml/src/ggml-cpu/repack.cpp
@@ -206,8 +206,9 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     const int ncols_interleaved = 4;
     const int blocklen = 4;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
 
     UNUSED(s);
     UNUSED(bs);
@@ -307,30 +308,28 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-    {
-        float sumf[8];
-        int sumi;
+    float sumf[8];
+    int sumi;
 
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
 
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
         }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
@@ -494,43 +493,73 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     const int ncols_interleaved = 4;
     const int blocklen = 4;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
     UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
-    {
-        float sumf[4];
-        int sumi;
+    float sumf[4];
+    int sumi;
 
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
 
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
                     }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
         }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
@@ -934,6 +963,50 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
+void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
 } // extern "C"
 
 static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
@@ -1285,15 +1358,16 @@ static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_s
 
 static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
-    //GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
     GGML_ASSERT(interleave_block == 4);
 
-    block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
-    const block_iq4_nl * src = (const block_iq4_nl *)data;
+    const block_iq4_nl   * src = (const block_iq4_nl   *)data;
+          block_iq4_nlx4 * dst = (      block_iq4_nlx4 *)t->data;
+
     block_iq4_nl dst_tmp[4];
+
     int nrow = ggml_nrows(t);
     int nrows_interleaved = 4;
-    int nblocks = t->ne[0] / QK4_0;
+    int nblocks = t->ne[0] / QK4_NL;
 
     GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
 
@@ -1315,6 +1389,63 @@ static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_b
     GGML_UNUSED(data_size);
 }
 
+static block_iq4_nlx8 make_block_iq4_nlx8(block_iq4_nl * in, unsigned int blck_size_interleave) {
+    block_iq4_nlx8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_NL * 4 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 8;
+            int src_offset = (i / 8) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
+    GGML_ASSERT(interleave_block == 8);
+
+    const block_iq4_nl   * src = (const block_iq4_nl   *)data;
+          block_iq4_nlx8 * dst = (      block_iq4_nlx8 *)t->data;
+
+    block_iq4_nl dst_tmp[8];
+
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 8;
+    int nblocks = t->ne[0] / QK4_NL;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
+
+    if (t->ne[1] % nrows_interleaved != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_iq4_nlx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
 namespace ggml::cpu::repack {
 // repack
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
@@ -1350,6 +1481,10 @@ template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void *
 //    return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
 //}
 
+template <> int repack<block_iq4_nl, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size);
+}
+
 // gemv
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemv(int, float *, size_t, const void *, const void *, int, int);
@@ -1378,6 +1513,10 @@ template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 // gemm
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemm(int, float *, size_t, const void *, const void *, int, int);
@@ -1406,6 +1545,10 @@ template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 class tensor_traits_base : public ggml::cpu::tensor_traits {
   public:
     virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
@@ -1680,6 +1823,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
 
     // instance for IQ4
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0;
 
     if (cur->type == GGML_TYPE_Q4_0) {
         if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
@@ -1710,6 +1854,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
             }
         }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
+        if (ggml_cpu_has_avx2()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &iq4_nl_8x8_q8_0;
+            }
+        }
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {
                 return &iq4_nl_4x4_q8_0;
diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h
index cd322e74389d6..cb32b503d3a11 100644
--- a/ggml/src/ggml-cpu/repack.h
+++ b/ggml/src/ggml-cpu/repack.h
@@ -67,6 +67,13 @@ struct block_iq4_nlx4 {
 
 static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
 
+struct block_iq4_nlx8 {
+    ggml_half d[8];            // deltas for 8 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 4];  // nibbles / quants for 8 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
+
 #if defined(__cplusplus)
 extern "C" {
 #endif
@@ -80,12 +87,14 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -97,12 +106,14 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 #if defined(__cplusplus)
 } // extern "C"
diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index 98ed29bc9c12f..ea824965aae2d 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -24,12 +24,6 @@ if (CUDAToolkit_FOUND)
         #     for best performance and to also build real architectures for the most commonly used GPUs.
         if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
             set(CMAKE_CUDA_ARCHITECTURES "native")
-        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
-                set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
-            else()
-                set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real")
-            endif()
         else()
             if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
                 set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
@@ -91,10 +85,6 @@ if (CUDAToolkit_FOUND)
         add_compile_definitions(GGML_CUDA_NO_FA)
     endif()
 
-    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-        add_compile_definitions(GGML_CUDA_F16)
-    endif()
-
     if (GGML_CUDA_NO_PEER_COPY)
         add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
     endif()
@@ -120,6 +110,10 @@ if (CUDAToolkit_FOUND)
 
     set(CUDA_FLAGS -use_fast_math -extended-lambda)
 
+    if (GGML_CUDA_DEBUG)
+        list(APPEND CUDA_FLAGS -lineinfo)
+    endif()
+
     if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
         # Options are:
         # - none (not recommended)
diff --git a/ggml/src/ggml-cuda/add-id.cu b/ggml/src/ggml-cuda/add-id.cu
index 8bed62ac9d215..8d9cf692b4b55 100644
--- a/ggml/src/ggml-cuda/add-id.cu
+++ b/ggml/src/ggml-cuda/add-id.cu
@@ -11,14 +11,14 @@ static __global__ void add_id_kernel(
     const int64_t i1 = blockIdx.x;
     const int64_t i2 = blockIdx.y;
 
-    const int i11 = *(int32_t *) ((char *) src2 + i1*sizeof(int32_t) + i2*nb21);
+    const int i11 = *(const int32_t *) ((const char *) src2 + i1*sizeof(int32_t) + i2*nb21);
 
     const size_t nb1 = ne0 * sizeof(float);
     const size_t nb2 = ne1 * nb1;
 
     float * dst_row = (float *)((char *)dst + i1*nb1 + i2*nb2);
-    const float * src0_row = (const float *)((char *)src0 +  i1*nb01 + i2*nb02);
-    const float * src1_row = (const float *)((char *)src1 + i11*nb11);
+    const float * src0_row = (const float *)((const char *)src0 +  i1*nb01 + i2*nb02);
+    const float * src1_row = (const float *)((const char *)src1 + i11*nb11);
 
     for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
         dst_row[i0] = src0_row[i0] + src1_row[i0];
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index 2e5d48797fa49..767ad83f60eb5 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -78,6 +78,8 @@
 #define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
+#define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons
+
 #define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
 #define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
 #define GGML_CUDA_CC_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
@@ -87,6 +89,10 @@
 #define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
 #define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+#    define GGML_CUDA_USE_CUB
+#endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+
 #ifdef __CUDA_ARCH_LIST__
 constexpr bool ggml_cuda_has_arch_impl(int) {
     return false;
@@ -200,14 +206,6 @@ static const char * cu_get_error_str(CUresult err) {
 #define GGML_CUDA_ASSUME(x)
 #endif // CUDART_VERSION >= 11010
 
-#ifdef GGML_CUDA_F16
-typedef half dfloat; // dequantize float
-typedef half2 dfloat2;
-#else
-typedef float dfloat; // dequantize float
-typedef float2 dfloat2;
-#endif // GGML_CUDA_F16
-
 #if (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
 #define GGML_USE_VMM
 #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
@@ -312,11 +310,11 @@ static bool turing_mma_available(const int cc) {
 }
 
 static bool ampere_mma_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
 }
 
 static bool cp_async_available(const int cc) {
-    return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE;
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
@@ -420,26 +418,6 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
-// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
-template<bool norm>
-static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.x;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col != 0) {
-        return;
-    }
-
-    dst[row] = norm ? sum / ncols : sum;
-}
-
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_all(int x) {
 #ifdef GGML_USE_HIP
@@ -480,25 +458,21 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
 }
 
 static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if defined(GGML_USE_HIP) && HIP_VERSION >= 50700000
+#if defined(GGML_USE_HIP)
     return half2(__hmax(a.x, b.x), __hmax(a.y, b.y));
-#elif !defined(GGML_USE_HIP) && CUDART_VERSION >= CUDART_HMAX
+#elif CUDART_VERSION >= CUDART_HMAX
     return __hmax2(a, b);
-#elif !defined(GGML_USE_HIP)
+#else
     half2 ret;
     reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
     reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
     return ret;
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
 #endif
 }
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 #pragma unroll
    for (int offset = width/2; offset > 0; offset >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, width));
@@ -507,16 +481,17 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 }
 
-#if CUDART_VERSION < CUDART_HMASK
+#if (defined(CUDART_VERSION) && CUDART_VERSION < CUDART_HMASK) || defined(GGML_USE_HIP) || \
+    (defined(MUSART_VERSION) && MUSART_VERSION < MUSART_HMASK)
 static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
     const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
     const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
     return mask_low | mask_high;
 }
-#endif // CUDART_VERSION < CUDART_HMASK
+#endif // (defined(CUDART_VERSION) && CUDART_VERSION < CUDART_HMASK) || defined(GGML_USE_HIP) || (defined(MUSART_VERSION) && MUSART_VERSION < MUSART_HMASK)
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
 #if defined(GGML_USE_HIP)
@@ -576,7 +551,7 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 #endif // CUDART_VERSION >= 12050
 }
 
-typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
+typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, float2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(
     const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
diff --git a/ggml/src/ggml-cuda/conv-transpose-1d.cu b/ggml/src/ggml-cuda/conv-transpose-1d.cu
index fe4caf674d4d9..8418ba667318b 100644
--- a/ggml/src/ggml-cuda/conv-transpose-1d.cu
+++ b/ggml/src/ggml-cuda/conv-transpose-1d.cu
@@ -34,10 +34,7 @@ static  __global__ void conv_transpose_1d_kernel(
         }
     }
     dst[global_index] = accumulator;
-    GGML_UNUSED(p0); GGML_UNUSED(d0); GGML_UNUSED(src0_ne3);
-    GGML_UNUSED(src1_ne3); GGML_UNUSED(dst_ne3);
-    GGML_UNUSED(src1_ne1); GGML_UNUSED(dst_ne1);
-    GGML_UNUSED(src1_ne2); GGML_UNUSED(dst_ne2);
+    GGML_UNUSED_VARS(p0, d0, src0_ne3, src1_ne3, dst_ne3, src1_ne1, dst_ne1, src1_ne2, dst_ne2);
 }
 
 static void conv_transpose_1d_f32_f32_cuda(
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
index e3beddbc1b23b..ba3d4eeb88085 100644
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -27,12 +27,12 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     const int64_t y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
-    dfloat2 v;
+    float2 v;
     dequantize_kernel(vx, ib, iqs, v);
 
     const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
-    y[iy0 + 0]        = float(v.x);
-    y[iy0 + y_offset] = float(v.y);
+    y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
+    y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
 }
 
 template <bool need_check>
@@ -71,9 +71,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
         y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d));
     }
 #else
-    GGML_UNUSED(vx);
-    GGML_UNUSED(y);
-    GGML_UNUSED(k);
+    GGML_UNUSED_VARS(vx, y, k);
     NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 }
@@ -630,7 +628,7 @@ static __global__ void convert_unary(
 
     const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
     const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
-    y[iy] = float(x[ix]);
+    y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
 }
 
 template <typename src_t, typename dst_t>
diff --git a/ggml/src/ggml-cuda/convert.cuh b/ggml/src/ggml-cuda/convert.cuh
index f04214be175ba..c62e8a1b1040a 100644
--- a/ggml/src/ggml-cuda/convert.cuh
+++ b/ggml/src/ggml-cuda/convert.cuh
@@ -29,3 +29,16 @@ typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
 to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
 to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
 to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);
+
+template<typename dst_t, typename src_t>
+ __host__ __device__ inline dst_t ggml_cuda_cast(src_t x) {
+    if constexpr (std::is_same_v<dst_t, src_t>) {
+        return x;
+    } else if constexpr(std::is_same_v<dst_t, nv_bfloat16>) {
+        return __float2bfloat16(float(x));
+    } else if constexpr(std::is_same_v<src_t, nv_bfloat16>) {
+        return __bfloat162float(x);
+    } else {
+        return float(x);
+    }
+}
diff --git a/ggml/src/ggml-cuda/cpy-utils.cuh b/ggml/src/ggml-cuda/cpy-utils.cuh
index 410c12b7ba56b..e621cb9811ab6 100644
--- a/ggml/src/ggml-cuda/cpy-utils.cuh
+++ b/ggml/src/ggml-cuda/cpy-utils.cuh
@@ -1,15 +1,7 @@
 #pragma once
 
 #include "ggml-common.h"
-
-template<typename src_t, typename dst_t>
-static __device__ __forceinline__ void convert_flt(const src_t * src, dst_t * dst) {
-    if constexpr (std::is_same_v<src_t, dst_t>) {
-        *dst = *src;
-    } else {
-        *dst = float(*src);
-    }
-}
+#include "convert.cuh"
 
 static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
     if (x <= val[0]) return 0;
@@ -221,5 +213,5 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
 
 template<typename src_t, typename dst_t>
 static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
-    convert_flt((const src_t *)cxi, (dst_t *)cdsti);
+    *(dst_t *) cdsti = ggml_cuda_cast<dst_t>(*(const src_t *) cxi);
 }
diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
index f9bb025643ca2..c40db08cedb05 100644
--- a/ggml/src/ggml-cuda/cpy.cu
+++ b/ggml/src/ggml-cuda/cpy.cu
@@ -42,7 +42,7 @@ static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
 
 #pragma unroll
     for (int j = 0; j < QK8_0; j += 2) {
-        dfloat2 dq;
+        float2 dq;
         dequantize_q8_0(cxi, 0, j, dq);
         *(cdstf + j) = dq.x;
         *(cdstf + j + 1) = dq.y;
@@ -55,7 +55,7 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
 
 #pragma unroll
     for (int j = 0; j < qk/2; j++) {
-        dfloat2 dq;
+        float2 dq;
         dequant(cxi, 0, j, dq);
         *(cdstf + j) = dq.x;
         *(cdstf + j + qk/2) = dq.y;
@@ -134,8 +134,7 @@ void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_des
     CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream));
     cuda_graph->graph_cpynode_index = 0; // reset index
 #else
-    GGML_UNUSED(cuda_graph); GGML_UNUSED(host_dest_ptrs);
-    GGML_UNUSED(host_dest_ptrs_size); GGML_UNUSED(stream);
+    GGML_UNUSED_VARS(cuda_graph, host_dest_ptrs, host_dest_ptrs_size, stream);
 #endif
 }
 
diff --git a/ggml/src/ggml-cuda/dequantize.cuh b/ggml/src/ggml-cuda/dequantize.cuh
index bd3c2d9db9463..e060fb29fdc03 100644
--- a/ggml/src/ggml-cuda/dequantize.cuh
+++ b/ggml/src/ggml-cuda/dequantize.cuh
@@ -1,48 +1,37 @@
 #include "common.cuh"
 
-static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q4_0 * x = (const block_q4_0 *) vx;
 
-    const dfloat d = x[ib].d;
+    const float d = x[ib].d;
 
     const int vui = x[ib].qs[iqs];
 
     v.x = vui & 0xF;
     v.y = vui >> 4;
 
-#ifdef GGML_CUDA_F16
-    v = __hsub2(v, {8.0f, 8.0f});
-    v = __hmul2(v, {d, d});
-#else
     v.x = (v.x - 8.0f) * d;
     v.y = (v.y - 8.0f) * d;
-#endif // GGML_CUDA_F16
 }
 
-static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q4_1 * x = (const block_q4_1 *) vx;
 
-    const dfloat d = __low2half(x[ib].dm);
-    const dfloat m = __high2half(x[ib].dm);
+    const float2 dm = __half22float2(x[ib].dm);
 
     const int vui = x[ib].qs[iqs];
 
     v.x = vui & 0xF;
     v.y = vui >> 4;
 
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-    v = __hadd2(v, {m, m});
-#else
-    v.x = (v.x * d) + m;
-    v.y = (v.y * d) + m;
-#endif // GGML_CUDA_F16
+    v.x = (v.x * dm.x) + dm.y;
+    v.y = (v.y * dm.x) + dm.y;
 }
 
-static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q5_0 * x = (const block_q5_0 *) vx;
 
-    const dfloat d = x[ib].d;
+    const float d = x[ib].d;
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
@@ -53,20 +42,14 @@ static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const in
     v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
     v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-#ifdef GGML_CUDA_F16
-    v = __hsub2(v, {16.0f, 16.0f});
-    v = __hmul2(v, {d, d});
-#else
     v.x = (v.x - 16.0f) * d;
     v.y = (v.y - 16.0f) * d;
-#endif // GGML_CUDA_F16
 }
 
-static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q5_1 * x = (const block_q5_1 *) vx;
 
-    const dfloat d = __low2half(x[ib].dm);
-    const dfloat m = __high2half(x[ib].dm);
+    const float2 dm = __half22float2(x[ib].dm);
 
     uint32_t qh;
     memcpy(&qh, x[ib].qh, sizeof(qh));
@@ -77,27 +60,18 @@ static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const in
     v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
     v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-    v = __hadd2(v, {m, m});
-#else
-    v.x = (v.x * d) + m;
-    v.y = (v.y * d) + m;
-#endif // GGML_CUDA_F16
+    v.x = (v.x * dm.x) + dm.y;
+    v.y = (v.y * dm.x) + dm.y;
 }
 
-static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q8_0 * x = (const block_q8_0 *) vx;
 
-    const dfloat d = x[ib].d;
+    const float d = x[ib].d;
 
     v.x = x[ib].qs[iqs + 0];
     v.y = x[ib].qs[iqs + 1];
 
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-#else
     v.x *= d;
     v.y *= d;
-#endif // GGML_CUDA_F16
 }
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index e46f0e2081bdf..b69f57d659a26 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -539,11 +539,15 @@ static __global__ void flash_attn_mask_to_KV_max(
         all_inf = warp_reduce_all(all_inf);
 
         if (!all_inf) {
-            KV_max_sj += FATTN_KQ_STRIDE;
             break;
         }
     }
 
+    // If the break in the loop was not triggered, KV_max_sj is now -FATTN_KQ_STRIDE.
+    // If the break was triggered it's the lower edge of the tile with the first non-masked values.
+    // In either case, walk back the decrementation by FATTN_KQ_STRIDE.
+    KV_max_sj += FATTN_KQ_STRIDE;
+
     if (threadIdx.x != 0) {
         return;
     }
@@ -700,28 +704,6 @@ static __global__ void flash_attn_combine_results(
     dst[tid] = VKQ_numerator / VKQ_denominator;
 }
 
-[[noreturn]]
-static void on_no_fattn_vec_case(const int D) {
-    if (D == 64) {
-        fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");
-        fprintf(stderr, "By default only f16 KV cache is supported.\n");
-        fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for V cache quantization support.\n");
-        GGML_ABORT("fatal error");
-    } else if (D == 128) {
-        fprintf(stderr, "Unsupported KV type combination for head_size 128.\n");
-        fprintf(stderr, "Supported combinations:\n");
-        fprintf(stderr, "  - K == q4_0, V == q4_0,  4.50 BPV\n");
-        fprintf(stderr, "  - K == q8_0, V == q8_0,  8.50 BPV\n");
-        fprintf(stderr, "  - K == f16,  V == f16,  16.00 BPV\n");
-        fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
-        GGML_ABORT("fatal error");
-    } else {
-        fprintf(stderr, "Unsupported KV type combination for head_size %d.\n", D);
-        fprintf(stderr, "Only f16 is supported.\n");
-        GGML_ABORT("fatal error");
-    }
-}
-
 template <int DV, int ncols1, int ncols2>
 void launch_fattn(
     ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index 39731baaeb7f4..57defb0c629d6 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -767,14 +767,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         }
     }
 #else
-    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
-    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
-    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V);
-    GGML_UNUSED(stride_mask); GGML_UNUSED(tile_K);
-    GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B);
-    GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum);
-    GGML_UNUSED(kb0); GGML_UNUSED(tile_Q);
+    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup,
+        scale, slope, logit_softcap, ne01, ne02,
+        stride_K, stride_V, stride_mask,
+        tile_Q, tile_K, tile_V, tile_mask,
+        Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
 }
@@ -1236,12 +1233,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         }
     }
 #else
-    GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2);
-    GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup);
-    GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_Q1);
-    GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V); GGML_UNUSED(stride_mask);
-    GGML_UNUSED(jt); GGML_UNUSED(kb0_start); GGML_UNUSED(kb0_stop);
+    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dstk_fixup,
+        scale, slope, logit_softcap, ne01, ne02,
+        stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
+        jt, kb0_start, kb0_stop);
     NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
 }
@@ -1395,17 +1390,15 @@ static __global__ void flash_attn_ext_f16(
         (Q_f2, K_h2, V_h2, mask_h2, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
          ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
-    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(TURING_MMA_AVAILABLE)
 }
diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index 1e23f8f79c202..a900799a991c0 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -258,7 +258,7 @@ static __global__ void flash_attn_tile_ext_f16(
             const half val = hexp(sink - kqmax[j0/nwarps]);
             kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
             if (threadIdx.x == 0) {
-                kqsum[j0/nwarps].x = __hadd(kqsum[j0/nwarps].x, val);
+                kqsum[j0/nwarps].x = __hadd(__low2half(kqsum[j0/nwarps]), val);
             }
 
 #pragma unroll
@@ -299,17 +299,15 @@ static __global__ void flash_attn_tile_ext_f16(
         }
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
-    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
-    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-    GGML_UNUSED(nb23);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index c58194937d7a6..b96a9ef971970 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -38,17 +38,15 @@ static __global__ void flash_attn_tile_ext_f32(
     return;
 #endif // FP16_MMA_AVAILABLE
     if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-        GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
-        GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-        GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-        GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-        GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-        GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-        GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-        GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+            max_bias, m0, m1, n_head_log2, logit_softcap,
+            ne00, ne01, ne02, ne03,
+                  nb01, nb02, nb03,
+            ne10, ne11, ne12, ne13,
+                  nb11, nb12, nb13,
+                  nb21, nb22, nb23,
+                  ne31, ne32, ne33,
+                  nb31, nb32, nb33);
         NO_DEVICE_CODE;
         return;
     }
@@ -312,17 +310,15 @@ static __global__ void flash_attn_tile_ext_f32(
         }
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
-    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }
diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
index b05f682cd3b4d..27a2dd6ae448f 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -349,17 +349,15 @@ static __global__ void flash_attn_vec_ext_f16(
         dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta);
-    GGML_UNUSED(scale); GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
index d6d0bfb744b74..da195d0334d77 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -37,17 +37,15 @@ static __global__ void flash_attn_vec_ext_f32(
 
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-        GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-        GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-        GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-        GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
-        GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-        GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
-        GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-        GGML_UNUSED(nb23);
+        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+            max_bias, m0, m1, n_head_log2, logit_softcap,
+            ne00, ne01, ne02, ne03,
+                  nb01, nb02, nb03,
+            ne10, ne11, ne12, ne13,
+                  nb11, nb12, nb13,
+                  nb21, nb22, nb23,
+                  ne31, ne32, ne33,
+                  nb31, nb32, nb33);
         NO_DEVICE_CODE;
         return;
     }
@@ -345,17 +343,15 @@ static __global__ void flash_attn_vec_ext_f32(
         dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
-    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }
diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
index fdc4d17da2da9..2219191fd9152 100644
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
@@ -15,7 +15,6 @@ namespace wmma = mtmusa::wmma;
 namespace wmma = nvcuda::wmma;
 #endif // GGML_USE_MUSA
 #elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
-#undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
 #include <rocwmma/rocwmma.hpp>
 namespace wmma = rocwmma;
 #endif // !defined(GGML_USE_HIP)
@@ -472,16 +471,15 @@ static __global__ void flash_attn_ext_f16(
         dst_meta[j_dst_unrolled] = dst_meta_val;
     }
 #else
-    GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(sinks);
-    GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
-    GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
-    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33); GGML_UNUSED(nb31);
-    GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
-    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
     NO_DEVICE_CODE;
 #endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)))
 }
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 22e90d0e7b316..48834272660e5 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -190,7 +190,7 @@ static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, gg
     FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
 #endif // GGML_CUDA_FA_ALL_QUANTS
 
-    on_no_fattn_vec_case(Q->ne[0]);
+    GGML_ABORT("fatal error");
 }
 
 #define FATTN_VEC_F32_CASE(D, type_K, type_V)                               \
@@ -265,74 +265,184 @@ static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, gg
     FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
 #endif // GGML_CUDA_FA_ALL_QUANTS
 
-    on_no_fattn_vec_case(Q->ne[0]);
+    GGML_ABORT("fatal error");
 }
 
-void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+// Best FlashAttention kernel for a specific GPU:
+enum best_fattn_kernel {
+    BEST_FATTN_KERNEL_NONE     =   0,
+    BEST_FATTN_KERNEL_TILE_F32 = 200,
+    BEST_FATTN_KERNEL_TILE_F16 = 210,
+    BEST_FATTN_KERNEL_VEC_F32  = 100,
+    BEST_FATTN_KERNEL_VEC_F16  = 110,
+    BEST_FATTN_KERNEL_WMMA_F16 = 300,
+    BEST_FATTN_KERNEL_MMA_F16  = 400,
+};
+
+static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const ggml_tensor * dst) {
+#ifndef FLASH_ATTN_AVAILABLE
+    GGML_UNUSED(device); GGML_UNUSED(dst);
+    return BEST_FATTN_KERNEL_NONE;
+#endif// FLASH_ATTN_AVAILABLE
+
     const ggml_tensor * KQV   = dst;
     const ggml_tensor * Q     = dst->src[0];
     const ggml_tensor * K     = dst->src[1];
     const ggml_tensor * V     = dst->src[2];
     const ggml_tensor * mask  = dst->src[3];
 
-    ggml_cuda_set_device(ctx.device);
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
+    const int gqa_ratio = Q->ne[2] / K->ne[2];
+    GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+
+    const int cc = ggml_cuda_info().devices[device].cc;
+    const int warp_size = ggml_cuda_info().devices[device].warp_size;
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
-#if defined(GGML_HIP_ROCWMMA_FATTN)
-    if (GGML_CUDA_CC_IS_AMD(cc) && fp16_mma_available(cc)) {
-        ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
-        return;
+    switch (K->ne[0]) {
+        case  64:
+        case 128:
+        case 256:
+            if (V->ne[0] != K->ne[0]) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
+        case  80:
+        case  96:
+        case 112:
+            if (V->ne[0] != K->ne[0]) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            if (!fp16_mma_available(cc) && !turing_mma_available(cc)) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
+        case 576:
+            if (V->ne[0] != 512) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            if (!turing_mma_available(cc) || gqa_ratio % 16 != 0) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
+        default:
+            return BEST_FATTN_KERNEL_NONE;
     }
-#endif // defined(GGML_HIP_ROCWMMA_FATTN)
 
-    if (!fast_fp16_available(cc)) {
-        if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
-            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-        } else {
-            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
-        }
-        return;
+#ifndef GGML_CUDA_FA_ALL_QUANTS
+    if (K->type != V->type) {
+        return BEST_FATTN_KERNEL_NONE;
     }
+#endif // GGML_CUDA_FA_ALL_QUANTS
 
-    if (!fp16_mma_available(cc)) {
-        if (prec == GGML_PREC_DEFAULT) {
-            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
-                ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-            } else {
-                ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+    switch (K->type) {
+        case GGML_TYPE_F16:
+            break;
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+#ifndef GGML_CUDA_FA_ALL_QUANTS
+            return BEST_FATTN_KERNEL_NONE;
+#endif // GGML_CUDA_FA_ALL_QUANTS
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+#ifdef GGML_CUDA_FA_ALL_QUANTS
+            if (K->ne[0] != 128 && K->ne[0] != 64) {
+                return BEST_FATTN_KERNEL_NONE;
             }
-        } else {
-            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
-                ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-            } else {
-                ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+#else
+            if (K->ne[0] != 128) {
+                return BEST_FATTN_KERNEL_NONE;
             }
-        }
-        return;
+#endif // GGML_CUDA_FA_ALL_QUANTS
+            break;
+        default:
+            return BEST_FATTN_KERNEL_NONE;
+    }
+
+    switch (V->type) {
+        case GGML_TYPE_F16:
+            break;
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            if (K->ne[0] != 128) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
+        default:
+            return BEST_FATTN_KERNEL_NONE;
+    }
+
+    if (mask && mask->ne[2] != 1) {
+        return BEST_FATTN_KERNEL_NONE;
     }
 
-    const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
-    const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
-    const bool mma_faster_for_rtx4000 = Q->ne[3] > 1 || (Q->ne[2] > 4*K->ne[2] && K->ne[1] >= 8192);
-    const bool mma_faster_for_bs1 = turing_mma_available(cc) && gqa_opt_applies && !mma_needs_data_conversion &&
-        (cc < GGML_CUDA_CC_ADA_LOVELACE || mma_faster_for_rtx4000);
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
-    if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
-        if (prec == GGML_PREC_DEFAULT) {
-            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-        } else {
-            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+
+    // If Turing tensor cores available, use them except for some cases with batch size 1:
+    if (turing_mma_available(cc)) {
+        const bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask; // The mma-based kernels have GQA-specific optimizations
+        const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
+        const bool mma_faster_for_rtx4000 = Q->ne[3] > 1 || (gqa_ratio > 4 && K->ne[1] >= 8192);
+        const bool mma_faster_for_bs1 = gqa_opt_applies && !mma_needs_data_conversion &&
+            (cc < GGML_CUDA_CC_ADA_LOVELACE || mma_faster_for_rtx4000);
+        if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
+            if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+                return BEST_FATTN_KERNEL_VEC_F16;
+            }
+            return BEST_FATTN_KERNEL_VEC_F32;
         }
-        return;
+        return BEST_FATTN_KERNEL_MMA_F16;
     }
 
-    // The MMA implementation needs Turing or newer, use the old WMMA code for Volta:
-    if (fp16_mma_available(cc) && !turing_mma_available(cc)) {
-        ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
-        return;
+    // Use kernels specializes for small batch sizes if possible:
+    if (Q->ne[1] <= 8 && can_use_vector_kernel) {
+        if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+            return BEST_FATTN_KERNEL_VEC_F16;
+        }
+        return BEST_FATTN_KERNEL_VEC_F32;
+    }
+
+    // For large batch sizes, use the WMMA kernel if possible:
+    if (fp16_mma_available(cc)) {
+        return BEST_FATTN_KERNEL_WMMA_F16;
+    }
+
+    // If there is no suitable kernel for tensor cores or small batch sizes, use the generic kernel for large batch sizes:
+    if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+        return BEST_FATTN_KERNEL_TILE_F16;
     }
+    return BEST_FATTN_KERNEL_TILE_F32;
+}
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_set_device(ctx.device);
+    switch (ggml_cuda_get_best_fattn_kernel(ggml_cuda_get_device(), dst)) {
+        case BEST_FATTN_KERNEL_NONE:
+            GGML_ABORT("fatal error");
+        case BEST_FATTN_KERNEL_TILE_F32:
+            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_TILE_F16:
+            ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_VEC_F32:
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_VEC_F16:
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_WMMA_F16:
+            ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+            break;
+        case BEST_FATTN_KERNEL_MMA_F16:
+            ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);
+            break;
+    }
+}
 
-    ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);
+bool ggml_cuda_flash_attn_ext_supported(int device, const ggml_tensor * dst) {
+    return ggml_cuda_get_best_fattn_kernel(device, dst) != BEST_FATTN_KERNEL_NONE;
 }
diff --git a/ggml/src/ggml-cuda/fattn.cuh b/ggml/src/ggml-cuda/fattn.cuh
index ad3ca7a8d8e4d..78705d59951c1 100644
--- a/ggml/src/ggml-cuda/fattn.cuh
+++ b/ggml/src/ggml-cuda/fattn.cuh
@@ -1,3 +1,5 @@
 #include "common.cuh"
 
 void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+bool ggml_cuda_flash_attn_ext_supported(int device, const ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/getrows.cu b/ggml/src/ggml-cuda/getrows.cu
index f77b2629a19b0..3ec0e957ab554 100644
--- a/ggml/src/ggml-cuda/getrows.cu
+++ b/ggml/src/ggml-cuda/getrows.cu
@@ -1,5 +1,6 @@
 #include "getrows.cuh"
 #include "dequantize.cuh"
+#include "convert.cuh"
 
 template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void k_get_rows(
@@ -31,11 +32,11 @@ static __global__ void k_get_rows(
     const int y_offset = qr == 1 ? 1 : qk/2;
 
     // dequantize
-    dfloat2 v;
+    float2 v;
     dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst_row[iybs + iqs + 0]        = float(v.x);
-    dst_row[iybs + iqs + y_offset] = float(v.y);
+    dst_row[iybs + iqs + 0]        = ggml_cuda_cast<dst_t>(v.x);
+    dst_row[iybs + iqs + y_offset] = ggml_cuda_cast<dst_t>(v.y);
 }
 
 template<typename src0_t, typename dst_t>
@@ -62,7 +63,7 @@ static __global__ void k_get_rows_float(
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
     const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
 
-    dst_row[i00] = float(src0_row[i00]);
+    dst_row[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
 }
 
 template<typename grad_t, typename dst_t>
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index d9110491ec78c..aa45ab39ed89e 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -28,6 +28,7 @@
 #include "ggml-cuda/mmvq.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
+#include "ggml-cuda/opt-step-sgd.cuh"
 #include "ggml-cuda/out-prod.cuh"
 #include "ggml-cuda/pad.cuh"
 #include "ggml-cuda/pool2d.cuh"
@@ -48,6 +49,7 @@
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
 #include "ggml-cuda/set-rows.cuh"
+#include "ggml-cuda/pad_reflect_1d.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -180,30 +182,6 @@ static int ggml_cuda_parse_id(char devName[]) {
 #endif // defined(GGML_USE_HIP)
 
 static ggml_cuda_device_info ggml_cuda_init() {
-#if defined(GGML_USE_HIP)
-    // Workaround for a rocBLAS bug when using multiple graphics cards:
-    // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
-    {
-        int major_version = 0;
-        size_t version_length = 0;
-        if (rocblas_get_version_string_size(&version_length) == rocblas_status_success) {
-            std::vector<char> version(version_length+1, '\0');
-            if (rocblas_get_version_string(version.data(), version.size()) == rocblas_status_success) {
-                version.resize(::strlen(version.data()));
-                int parsed_value = 0;
-                if (std::from_chars(version.data(), version.data() + version.size(), parsed_value).ec == std::errc()) {
-                    major_version = parsed_value;
-                }
-            }
-        }
-        if (major_version < 4) {
-            GGML_LOG_DEBUG(GGML_CUDA_NAME " calling rocblas_initialize as a workaround for a rocBLAS bug\n");
-            rocblas_initialize();
-            CUDA_CHECK(cudaDeviceSynchronize());
-        }
-    }
-#endif
-
     ggml_cuda_device_info info = {};
 
     cudaError_t err = cudaGetDeviceCount(&info.device_count);
@@ -1351,9 +1329,7 @@ static void ggml_cuda_op_mul_mat_cublas(
                     &beta,  dst_dd_i,    ldc));
     }
 
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddq_i);
-    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED_VARS(dst, src1_ddq_i, src1_padded_row_size);
 }
 
 static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
@@ -2377,6 +2353,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_PAD:
             ggml_cuda_op_pad(ctx, dst);
             break;
+        case GGML_OP_PAD_REFLECT_1D:
+            ggml_cuda_op_pad_reflect_1d(ctx, dst);
+            break;
         case GGML_OP_ARANGE:
             ggml_cuda_op_arange(ctx, dst);
             break;
@@ -2503,6 +2482,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_OPT_STEP_ADAMW:
             ggml_cuda_opt_step_adamw(ctx, dst);
             break;
+        case GGML_OP_OPT_STEP_SGD:
+            ggml_cuda_opt_step_sgd(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -3503,15 +3485,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
-        case GGML_OP_SUM_ROWS:
-        case GGML_OP_MEAN:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
             return true;
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
@@ -3519,47 +3502,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_GATED_LINEAR_ATTN:
         case GGML_OP_RWKV_WKV7:
             return true;
-        case GGML_OP_FLASH_ATTN_EXT: {
-#ifndef FLASH_ATTN_AVAILABLE
-            return false;
-#endif // FLASH_ATTN_AVAILABLE
-            if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
-                const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
-                if (!turing_mma_available(cc)) {
-                    return false;
-                }
-                const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];
-                return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
-            }
-            // TODO: more general-purpose attention sink support [TAG_ATTN_SINKS]
-            if (op->src[4] && !fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc)
-                    && op->src[0]->ne[0] != 64 && op->src[0]->ne[0] != 128) {
-                return false;
-            }
-            if (op->src[0]->ne[0] == 192) {
-                return false;
-            }
-            if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
-                return false;
-            }
-            if (op->src[0]->ne[0] ==  64 && op->src[1]->type == GGML_TYPE_F16) {
-                return true;
-            }
-            if (op->src[0]->ne[0] == 128) {
-                return true;
-            }
-            if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
-                return true;
-            }
-            if (op->src[3] && op->src[3]->ne[2] != 1) {
-                return false;
-            }
-            return fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc) &&
-                op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
-        }
+        case GGML_OP_FLASH_ATTN_EXT:
+            return ggml_cuda_flash_attn_ext_supported(dev_ctx->device, op);
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             return true;
         default:
             return false;
@@ -3691,10 +3639,6 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
         features.push_back({ "NO_PEER_COPY", "1" });
     #endif
 
-    #ifdef GGML_CUDA_F16
-        features.push_back({ "F16", "1" });
-    #endif
-
     #ifdef GGML_CUDA_USE_GRAPHS
         features.push_back({ "USE_GRAPHS", "1" });
     #endif
diff --git a/ggml/src/ggml-cuda/mean.cu b/ggml/src/ggml-cuda/mean.cu
index 4b238a3998ba3..347abc18660ca 100644
--- a/ggml/src/ggml-cuda/mean.cu
+++ b/ggml/src/ggml-cuda/mean.cu
@@ -1,4 +1,14 @@
 #include "mean.cuh"
+#include "reduce_rows.cuh"
+
+#ifdef GGML_CUDA_USE_CUB
+#include <cub/cub.cuh>
+using namespace cub;
+#endif  // GGML_CUDA_USE_CUB
+
+template <typename T> __global__ void divide_by_count(T * result, size_t count) {
+    *result /= static_cast<T>(count);
+}
 
 void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0   = dst->src[0];
@@ -13,7 +23,51 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    const dim3 block_dims(WARP_SIZE, 1, 1);
+// Special case for reducing vectors
+#ifdef GGML_CUDA_USE_CUB
+#ifdef USE_CUDA_GRAPH
+    cudaStreamCaptureStatus iscapturing;
+    CUDA_CHECK(cudaStreamIsCapturing(stream, &iscapturing));
+#endif // USE_CUDA_GRAPH
+    if ((nrows == 1) &&
+#ifdef USE_CUDA_GRAPH
+            // CUDA_GRAPHS_DISABLED
+            ((ncols > 65536) &&
+             ((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
+              ctx.cuda_graph->disable_due_to_gpu_arch || ctx.cuda_graph->disable_due_to_too_many_updates ||
+              ctx.cuda_graph->disable_due_to_failed_graph_capture)) ||
+        // CUDA_GRAPHS ENABLED
+        ((ncols > 32768) &&
+         !((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
+           ctx.cuda_graph->disable_due_to_gpu_arch || ctx.cuda_graph->disable_due_to_too_many_updates ||
+           ctx.cuda_graph->disable_due_to_failed_graph_capture))) {
+#else
+        (ncols > 65536)) {
+#endif // USE_CUDA_GRAPH
+        // Single row - use device-wide reduction
+        size_t           tmp_size = 0;
+        ggml_cuda_pool & pool     = ctx.pool();
+
+        DeviceReduce::Sum(nullptr, tmp_size, src0_d, dst_d, ncols, stream);
+
+        ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
+        DeviceReduce::Sum(tmp_alloc.ptr, tmp_size, src0_d, dst_d, ncols, stream);
+
+        // Divide by ncols
+        divide_by_count<float><<<1, 1, 0, stream>>>(dst_d, ncols);
+        return;
+    }
+#endif // GGML_CUDA_USE_CUB
+
     const dim3 block_nums(nrows, 1, 1);
-    reduce_rows_f32</*norm*/ true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+
+    const int id  = ggml_cuda_get_device();
+    const int nsm = ggml_cuda_info().devices[id].nsm;
+    if ((nrows / nsm) < 2) {
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    } else {
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    }
 }
diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh
index 83ee16b27d0df..667deb9c65017 100644
--- a/ggml/src/ggml-cuda/mma.cuh
+++ b/ggml/src/ggml-cuda/mma.cuh
@@ -291,9 +291,7 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[2]), "=r"(xi[1]), "=r"(xi[3])
             : "l"(xs));
 #else
-        GGML_UNUSED(t);
-        GGML_UNUSED(xs0);
-        GGML_UNUSED(stride);
+        GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -315,9 +313,7 @@ namespace ggml_cuda_mma {
             : "r"(A.x[1]), "r"(B.x[0]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -345,9 +341,7 @@ namespace ggml_cuda_mma {
             : "r"(A.x[3]), "r"(B.x[1]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -372,9 +366,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -408,9 +400,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -425,9 +415,7 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
             : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]));
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMPERE_MMA_AVAILABLE
     }
@@ -452,9 +440,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -469,9 +455,7 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
             : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]));
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMPERE_MMA_AVAILABLE
     }
@@ -505,9 +489,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
@@ -533,9 +515,7 @@ namespace ggml_cuda_mma {
                                                       0, 0, 0);
 #endif // defined(CDNA3)
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }
@@ -561,9 +541,7 @@ namespace ggml_cuda_mma {
                                                      0, 0, 0);
 #endif // defined(CDNA3)
 #else
-        GGML_UNUSED(D);
-        GGML_UNUSED(A);
-        GGML_UNUSED(B);
+        GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }
diff --git a/ggml/src/ggml-cuda/mmf.cu b/ggml/src/ggml-cuda/mmf.cu
index 1437367e87106..cfa5c5cce2c23 100644
--- a/ggml/src/ggml-cuda/mmf.cu
+++ b/ggml/src/ggml-cuda/mmf.cu
@@ -132,11 +132,11 @@ static __global__ void mul_mat_f(
         dst[j*stride_col_dst + row0 + threadIdx.x] = sum;
     }
 #else
+    GGML_UNUSED_VARS(x, y, ids, dst,
+        ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
+        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     NO_DEVICE_CODE;
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(ids); GGML_UNUSED(dst);
-    GGML_UNUSED(ncols); GGML_UNUSED(nchannels_y); GGML_UNUSED(stride_row); GGML_UNUSED(stride_col_y); GGML_UNUSED(stride_col_dst);
-    GGML_UNUSED(channel_ratio); GGML_UNUSED(stride_channel_x); GGML_UNUSED(stride_channel_y); GGML_UNUSED(stride_channel_dst);
-    GGML_UNUSED(sample_ratio); GGML_UNUSED(stride_sample_x); GGML_UNUSED(stride_sample_y); GGML_UNUSED(stride_sample_dst);
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
@@ -151,7 +151,6 @@ static void mul_mat_f_cuda(
         cudaStream_t stream) {
     typedef tile<16, 8, T>     tile_A;
     typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
 
     GGML_ASSERT(!ids && "mul_mat_id not implemented");
 
@@ -352,9 +351,6 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
     GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
     GGML_ASSERT(        nb0        == ts_dst);
 
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
-
     const float   * src1_d =       (const float   *) src1->data;
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
     float         *  dst_d =       (float         *)  dst->data;
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index 384ee7615f7a4..576032a0ce0dd 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -266,10 +266,7 @@ void ggml_cuda_op_mul_mat_q(
 
     ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddf_i);
-    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_padded_row_size);
 }
 
 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 96129bd831fd4..650f7080677ad 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -1255,7 +1255,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
+    GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
 }
@@ -1572,7 +1572,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
+    GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
 }
@@ -2301,7 +2301,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
         }
     }
 #else
-    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00);
+    GGML_UNUSED_VARS(x, y, sum, k00);
     NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
 }
@@ -2855,12 +2855,14 @@ static __device__ __forceinline__ void mmq_write_back_mma(
 #else
     typedef tile<16, 8, int> tile_C;
     constexpr int rows_per_warp = 2 * granularity;
-#endif
+#endif // defined(AMD_MFMA_AVAILABLE)
     constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
 
     const int i0 = (threadIdx.y / ntx) * (ntx*tile_C::I);
 #if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
     static_assert(nwarps*tile_C::I == mmq_y, "nwarps*tile_C::I != mmq_y");
+#else
+    GGML_UNUSED(nwarps);
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
 
 #pragma unroll
diff --git a/ggml/src/ggml-cuda/mmvf.cu b/ggml/src/ggml-cuda/mmvf.cu
index 1ad4bc75ba614..5b21ef05b3c35 100644
--- a/ggml/src/ggml-cuda/mmvf.cu
+++ b/ggml/src/ggml-cuda/mmvf.cu
@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "common.cuh"
+#include "convert.cuh"
 #include "mmvf.cuh"
 
 template <typename T, typename type_acc, int ncols_dst, int block_size>
@@ -93,8 +94,8 @@ static __global__ void mul_mat_vec_f(
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
-                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
-                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
+                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
             }
         }
     } else {
@@ -432,12 +433,7 @@ void ggml_cuda_op_mul_mat_vec_f(
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
     }
 
-    GGML_UNUSED(ctx);
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddq_i);
-    GGML_UNUSED(src1_ncols);
-    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED_VARS(ctx, src1, dst, src1_ddq_i, src1_ncols, src1_padded_row_size);
 }
 
 bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) {
diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index 5c8e5c4a7eeb5..b7c3079308e3f 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -596,9 +596,5 @@ void ggml_cuda_op_mul_mat_vec_q(
         src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
         1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
 
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddf_i);
-    GGML_UNUSED(src1_ncols);
-    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);
 }
diff --git a/ggml/src/ggml-cuda/opt-step-sgd.cu b/ggml/src/ggml-cuda/opt-step-sgd.cu
new file mode 100644
index 0000000000000..460b16de447af
--- /dev/null
+++ b/ggml/src/ggml-cuda/opt-step-sgd.cu
@@ -0,0 +1,49 @@
+#include "ggml-impl.h"
+#include "opt-step-sgd.cuh"
+
+#include <cstdint>
+
+static __global__ void opt_step_sgd_f32(
+    float * __restrict__ x, const float * __restrict__ g,
+    const float * __restrict__ pars, const int64_t k) {
+
+    const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    x[i] = x[i] * (1.0f - pars[0] * pars[1]) - pars[0] * g[i];
+}
+
+static void opt_step_sgd_f32_cuda(
+    float * x, const float * g, const float * __restrict__ pars, const int64_t k, cudaStream_t stream) {
+
+    const dim3 block_dims(CUDA_OPT_STEP_SGD_BLOCK_SIZE, 1, 1);
+    const dim3 block_nums((k + CUDA_OPT_STEP_SGD_BLOCK_SIZE - 1) / CUDA_OPT_STEP_SGD_BLOCK_SIZE, 1, 1);
+    opt_step_sgd_f32<<<block_nums, block_dims, 0, stream>>>(x, g, pars, k);
+}
+
+void ggml_cuda_opt_step_sgd(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0      = dst->src[0];
+    const ggml_tensor * src0_grad = dst->src[1];
+    const ggml_tensor * params    = dst->src[2];
+
+    GGML_ASSERT(src0->type      == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad->type == GGML_TYPE_F32);
+    GGML_ASSERT(params->type    == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src0_grad));
+    GGML_ASSERT(ggml_is_contiguous(params));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_nelements(params) == 2);
+
+    float       * src0_d      = (float       *) src0->data;
+    const float * src0_grad_d = (const float *) src0_grad->data;
+    const float * params_d    = (const float *) params->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t ne = ggml_nelements(src0);
+
+    opt_step_sgd_f32_cuda(src0_d, src0_grad_d, params_d, ne, stream);
+}
diff --git a/ggml/src/ggml-cuda/opt-step-sgd.cuh b/ggml/src/ggml-cuda/opt-step-sgd.cuh
new file mode 100644
index 0000000000000..f97ab7d9bede3
--- /dev/null
+++ b/ggml/src/ggml-cuda/opt-step-sgd.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_OPT_STEP_SGD_BLOCK_SIZE 256
+
+void ggml_cuda_opt_step_sgd(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/pad_reflect_1d.cu b/ggml/src/ggml-cuda/pad_reflect_1d.cu
new file mode 100644
index 0000000000000..4ed34aec3d331
--- /dev/null
+++ b/ggml/src/ggml-cuda/pad_reflect_1d.cu
@@ -0,0 +1,82 @@
+#include "pad_reflect_1d.cuh"
+
+static __global__ void pad_reflect_1d_kernel_f32(
+    const void * __restrict__ src0,
+    void * __restrict__ dst,
+    const int64_t ne0,
+    const int64_t ne00,
+    const int64_t ne01,
+    const int64_t ne02,
+    const int64_t ne03,
+    const int64_t nb00,
+    const int64_t nb01,
+    const int64_t nb02,
+    const int64_t nb03,
+    const int64_t nb0,
+    const int64_t nb1,
+    const int64_t nb2,
+    const int64_t nb3,
+    const int p0,
+    const int p1) {
+
+    const int64_t i3 = blockIdx.z;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
+
+    if (i1 >= ne01 || i2 >= ne02 || i3 >= ne03) {
+        return;
+    }
+
+    const char * src0_ptr = (const char *)src0 + i3*nb03 + i2*nb02 + i1*nb01;
+    char * dst_ptr = (char *)dst + i3*nb3 + i2*nb2 + i1*nb1;
+
+    for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+        float value;
+
+        if (i0 < p0) {
+            // Left padding - reflect
+            value = *(const float *)(src0_ptr + (p0 - i0) * nb00);
+        } else if (i0 < ne0 - p1) {
+            // Middle - copy
+            value = *(const float *)(src0_ptr + (i0 - p0) * nb00);
+        } else {
+            // Right padding - reflect
+            int64_t src_idx = (ne0 - p1 - p0) - (p1 + 1 - (ne0 - i0)) - 1;
+            value = *(const float *)(src0_ptr + src_idx * nb00);
+        }
+
+        *(float *)(dst_ptr + i0 * nb0) = value;
+    }
+}
+
+void ggml_cuda_op_pad_reflect_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *) dst->op_params;
+    const int p0 = opts[0];
+    const int p1 = opts[1];
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne0 = dst->ne[0];
+
+    GGML_ASSERT(ne0 == ne00 + p0 + p1);
+
+    const dim3 block_dims(CUDA_PAD_REFLECT_1D_BLOCK_SIZE, 1, 1);
+    const dim3 grid_dims(ne01, ne02, ne03);
+
+    pad_reflect_1d_kernel_f32<<<grid_dims, block_dims, 0, stream>>>(
+        src0->data, dst->data,
+        ne0, ne00, ne01, ne02, ne03,
+        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+        dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+        p0, p1
+    );
+}
diff --git a/ggml/src/ggml-cuda/pad_reflect_1d.cuh b/ggml/src/ggml-cuda/pad_reflect_1d.cuh
new file mode 100644
index 0000000000000..15f2ed1737b1a
--- /dev/null
+++ b/ggml/src/ggml-cuda/pad_reflect_1d.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_PAD_REFLECT_1D_BLOCK_SIZE 256
+
+void ggml_cuda_op_pad_reflect_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/reduce_rows.cuh b/ggml/src/ggml-cuda/reduce_rows.cuh
new file mode 100644
index 0000000000000..6bcae9e52fbee
--- /dev/null
+++ b/ggml/src/ggml-cuda/reduce_rows.cuh
@@ -0,0 +1,53 @@
+#include "common.cuh"
+
+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template <bool norm>
+static __global__ void reduce_rows_f32(const float * __restrict__ x, float * __restrict__ dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float     sum        = 0.0f;
+    const int num_unroll = 8;
+    float     temp[num_unroll];
+    float     sum_temp[num_unroll] = { 0.0f };
+    for (int i = col; i < ncols;) {
+        for (int j = 0; j < num_unroll; ++j) {
+            if (i < ncols) {
+                temp[j] = x[row * ncols + i];
+            } else {
+                temp[j] = 0;
+            }
+            i += blockDim.x;
+        }
+        for (int j = 0; j < num_unroll; ++j) {
+            sum_temp[j] += temp[j];
+        }
+    }
+    for (int j = 0; j < num_unroll; ++j) {
+        sum += sum_temp[j];
+    }
+
+    // sum up partial sums
+    sum = warp_reduce_sum(sum);
+    if (blockDim.x > WARP_SIZE) {
+        assert((blockDim.x <= 1024) && (blockDim.x % WARP_SIZE) == 0);
+        __shared__ float s_sum[32];
+        const int        warp_id = threadIdx.x / WARP_SIZE;
+        const int        lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = sum;
+        }
+        __syncthreads();
+        sum = 0.0f;
+        if (lane_id < (static_cast<int>(blockDim.x) / WARP_SIZE)) {
+            sum = s_sum[lane_id];
+        }
+        sum = warp_reduce_sum(sum);
+    }
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}
diff --git a/ggml/src/ggml-cuda/set-rows.cu b/ggml/src/ggml-cuda/set-rows.cu
index 07983436459d4..b4115a43c2a32 100644
--- a/ggml/src/ggml-cuda/set-rows.cu
+++ b/ggml/src/ggml-cuda/set-rows.cu
@@ -3,11 +3,6 @@
 
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);
 
-template<typename src_t, typename dst_t>
-__device__ __forceinline__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
-    convert_flt(src_f, dst_f);
-}
-
 // Generic quantized set_rows kernel template
 template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
 static __global__ void k_set_rows_quant(
@@ -117,9 +112,7 @@ static __global__ void k_set_rows(
     const src_t * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
     dst_t * dst_row_ptr    = dst + dst_row*s1 + i02*s2 + i03*s3;
 
-    const src_t* src_elem = src0_row + i00;
-    dst_t* dst_elem = dst_row_ptr + i00;
-    set_rows_1(src_elem, dst_elem);
+    dst_row_ptr[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
 
     GGML_UNUSED(ne10);
     GGML_UNUSED(ne13);
diff --git a/ggml/src/ggml-cuda/sum.cu b/ggml/src/ggml-cuda/sum.cu
index eb3d7cdba98a7..c56257b440661 100644
--- a/ggml/src/ggml-cuda/sum.cu
+++ b/ggml/src/ggml-cuda/sum.cu
@@ -1,19 +1,15 @@
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
-#define USE_CUB
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+#include "sum.cuh"
+#include "sumrows.cuh"
 
-#ifdef USE_CUB
+#ifdef GGML_CUDA_USE_CUB
 #include <cub/cub.cuh>
 using namespace cub;
-#endif // USE_CUB
-
-#include "sumrows.cuh"
-#include "sum.cuh"
+#endif  // GGML_CUDA_USE_CUB
 
 #include <cstdint>
 
 void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) {
-#ifdef USE_CUB
+#ifdef GGML_CUDA_USE_CUB
     size_t tmp_size = 0;
     DeviceReduce::Sum(nullptr,       tmp_size, x, dst, ne, stream);
     ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
@@ -23,7 +19,7 @@ void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int
     // For AMD there is rocPRIM which could be used as a drop-in replacement via hipcub but this would require C++11 -> C++14.
     sum_rows_f32_cuda(x, dst, ne, 1, stream);
     GGML_UNUSED(pool);
-#endif // USE_CUB
+#endif // GGML_CUDA_USE_CUB
 }
 
 void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
diff --git a/ggml/src/ggml-cuda/sumrows.cu b/ggml/src/ggml-cuda/sumrows.cu
index 2eee08fa07375..4025771aadb9d 100644
--- a/ggml/src/ggml-cuda/sumrows.cu
+++ b/ggml/src/ggml-cuda/sumrows.cu
@@ -1,9 +1,17 @@
+#include "reduce_rows.cuh"
 #include "sumrows.cuh"
 
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const int  id  = ggml_cuda_get_device();
+    const int  nsm = ggml_cuda_info().devices[id].nsm;
     const dim3 block_nums(nrows, 1, 1);
-    reduce_rows_f32</*norm*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    if ((nrows / nsm) < 2) {
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    } else {
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    }
 }
 
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -19,8 +27,17 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    const dim3 block_dims(WARP_SIZE, 1, 1);
     const dim3 block_nums(nrows, 1, 1);
 
-    reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    const int id  = ggml_cuda_get_device();
+    const int nsm = ggml_cuda_info().devices[id].nsm;
+    if ((nrows / nsm) < 2) {
+        // Increase num threads to 512 for small nrows to better hide the latency
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    } else {
+        // Enough active SMs to hide latency, use smaller blocks to allow better scheduling
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    }
 }
diff --git a/ggml/src/ggml-cuda/vecdotq.cuh b/ggml/src/ggml-cuda/vecdotq.cuh
index d8f9aa5ba6224..d60292b83b106 100644
--- a/ggml/src/ggml-cuda/vecdotq.cuh
+++ b/ggml/src/ggml-cuda/vecdotq.cuh
@@ -87,7 +87,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp
         sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi);
     }
 
-#ifdef GGML_CUDA_F16
+#ifdef FAST_FP16_AVAILABLE
     const float2 tmp = __half22float2(__hmul2(dm4, ds8));
     const float d4d8 = tmp.x;
     const float m4s8 = tmp.y;
@@ -96,7 +96,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp
     const float2 ds8f = __half22float2(ds8);
     const float d4d8 = dm4f.x * ds8f.x;
     const float m4s8 = dm4f.y * ds8f.y;
-#endif // GGML_CUDA_F16
+#endif // FAST_FP16_AVAILABLE
 
     // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
     return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
@@ -158,7 +158,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp
         sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
     }
 
-#ifdef GGML_CUDA_F16
+#ifdef FAST_FP16_AVAILABLE
     const float2 tmp = __half22float2(__hmul2(dm5, ds8));
     const float d5d8 = tmp.x;
     const float m5s8 = tmp.y;
@@ -167,7 +167,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp
     const float2 ds8f = __half22float2(ds8);
     const float d5d8 = dm5f.x * ds8f.x;
     const float m5s8 = dm5f.y * ds8f.y;
-#endif // GGML_CUDA_F16
+#endif // FAST_FP16_AVAILABLE
 
     // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
     return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
@@ -201,7 +201,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp
         sumi = ggml_cuda_dp4a(v[i], u[i], sumi);
     }
 
-#ifdef GGML_CUDA_F16
+#ifdef FAST_FP16_AVAILABLE
     const float2 tmp = __half22float2(__hmul2(dm8, ds8));
     const float d8d8 = tmp.x;
     const float m8s8 = tmp.y;
@@ -210,7 +210,7 @@ template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp
     const float2 ds8f = __half22float2(ds8);
     const float d8d8 = dm8f.x * ds8f.x;
     const float m8s8 = dm8f.y * ds8f.y;
-#endif // GGML_CUDA_F16
+#endif // FAST_FP16_AVAILABLE
 
     // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
     return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
index c31f319232252..6e9c67aca096e 100644
--- a/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -1,12 +1,10 @@
 #pragma once
 
-#define HIP_ENABLE_WARP_SYNC_BUILTINS 1
+#define HIP_DISABLE_WARP_SYNC_BUILTINS 1
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
-#include <hip/hip_bfloat16.h>
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
+#include <hip/hip_bf16.h>
 
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
@@ -137,7 +135,7 @@
 #define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
 #define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
 
-#if HIP_VERSION >= 70000000
+#if HIP_VERSION >= 60500000
 #define CUBLAS_COMPUTE_16F HIPBLAS_COMPUTE_16F
 #define CUBLAS_COMPUTE_32F HIPBLAS_COMPUTE_32F
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_COMPUTE_32F_FAST_16F
@@ -149,7 +147,7 @@
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define cublasComputeType_t hipblasDatatype_t
 #define cudaDataType_t hipblasDatatype_t
-#endif // HIP_VERSION >= 7000000
+#endif // HIP_VERSION >= 6050000
 
 #if !defined(__HIP_PLATFORM_AMD__)
 #error "The HIP backend supports only AMD targets"
@@ -181,8 +179,7 @@
 #define RDNA4
 #endif
 
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
-    defined(__gfx1150__) || defined(__gfx1151__)
+#if defined(__GFX11__)
 #define RDNA3
 #endif
 
@@ -199,8 +196,8 @@
     #define __has_builtin(x) 0
 #endif
 
-typedef hip_bfloat16 nv_bfloat16;
-typedef short2 nv_bfloat162; // FIXME there is no 2x BF16 type being defined in bfloat16.h, ad-hoc compilation fix
+typedef __hip_bfloat16 nv_bfloat16;
+typedef __hip_bfloat162 nv_bfloat162;
 
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
 typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
@@ -251,17 +248,3 @@ static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigne
     }
     return c;
 }
-
-#if HIP_VERSION < 50600000
-// __shfl_xor() for half2 was added in ROCm 5.6
-static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
-    typedef union half2_b32 {
-        half2 val;
-        int   b32;
-    } half2_b32_t;
-    half2_b32_t tmp;
-    tmp.val = var;
-    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
-    return tmp.val;
-}
-#endif // HIP_VERSION < 50600000
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index 852de97346025..d327b90cceb25 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -46,8 +46,8 @@ if (GGML_HIP_ROCWMMA_FATTN)
     endif()
 endif()
 
-if (${hip_VERSION} VERSION_LESS 5.5)
-    message(FATAL_ERROR "At least ROCM/HIP V5.5 is required")
+if (${hip_VERSION} VERSION_LESS 6.1)
+    message(FATAL_ERROR "At least ROCM/HIP V6.1 is required")
 endif()
 
 message(STATUS "HIP and hipBLAS found")
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index cb8eff4a77292..7c70d352dfddf 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -1846,7 +1846,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_ROPE:
             return true;
         case GGML_OP_IM2COL:
-            return op->src[0]->type == GGML_TYPE_F16;
+            return op->src[1]->type == GGML_TYPE_F32 && (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
         case GGML_OP_POOL_1D:
             return false;
         case GGML_OP_UPSCALE:
@@ -4703,7 +4703,6 @@ static int ggml_metal_encode_node(
             {
                 GGML_ASSERT(ggml_is_contiguous(src0));
                 GGML_ASSERT(ggml_is_contiguous(src1));
-                GGML_ASSERT(src0->type == GGML_TYPE_F16);
                 GGML_ASSERT(src1->type == GGML_TYPE_F32);
                 GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
 
diff --git a/ggml/src/ggml-musa/CMakeLists.txt b/ggml/src/ggml-musa/CMakeLists.txt
index 02904526ade04..cdb3818c786c7 100644
--- a/ggml/src/ggml-musa/CMakeLists.txt
+++ b/ggml/src/ggml-musa/CMakeLists.txt
@@ -96,10 +96,6 @@ if (MUSAToolkit_FOUND)
         add_compile_definitions(GGML_CUDA_NO_FA)
     endif()
 
-    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-        add_compile_definitions(GGML_CUDA_F16)
-    endif()
-
     if (GGML_CUDA_NO_PEER_COPY)
         add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
     endif()
diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt
index d8290faa467d0..9a7ccbcff0605 100644
--- a/ggml/src/ggml-opencl/CMakeLists.txt
+++ b/ggml/src/ggml-opencl/CMakeLists.txt
@@ -82,7 +82,9 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
     mul_mv_q6_k
+    mul_mv_mxfp4_f32
     mul_mv_id_q4_0_f32_8x_flat
+    mul_mv_id_mxfp4_f32
     mul_mm_f32_f32_l4_lm
     mul_mm_f16_f32_l4_lm
     mul
@@ -110,6 +112,9 @@ set(GGML_OPENCL_KERNELS
     mul_mat_f16_f32
     conv2d
     conv2d_f16_f32
+    flash_attn_f32_f16
+    flash_attn_f16
+    flash_attn_f32
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 8ba1e00df7bad..df27501361f7f 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -25,6 +25,7 @@
 #include <vector>
 #include <string>
 #include <cmath>
+#include <map>
 #include <memory>
 #include <charconv>
 #include <mutex>
@@ -332,6 +333,7 @@ struct ggml_backend_opencl_context {
 
     cl_int alignment;
     size_t max_alloc_size;
+    size_t max_workgroup_size;
     bool fp16_support;
     bool has_vector_subgroup_broadcast;
     bool disable_fusion;
@@ -365,6 +367,7 @@ struct ggml_backend_opencl_context {
     cl_program program_mul_mv_q4_0_f32_1d_8x_flat;
     cl_program program_mul_mv_q4_0_f32_1d_16x_flat;
     cl_program program_mul_mv_q6_K;
+    cl_program program_mul_mv_mxfp4_f32;
     cl_program program_mul_mv_f16_f16;
     cl_program program_mul_mv_f16_f32_1row;
     cl_program program_mul_mv_f16_f32_l4;
@@ -398,6 +401,7 @@ struct ggml_backend_opencl_context {
     cl_program program_conv_2d_f16_f32;
     cl_program program_tsembd;
     cl_program program_mul_mv_id_q4_0_f32_8x_flat;
+    cl_program program_mul_mv_id_mxfp4_f32;
     cl_program program_mul_mm_f32_f32_l4_lm;
     cl_program program_mul_mm_f16_f32_l4_lm;
 
@@ -422,6 +426,14 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8;
     cl_kernel kernel_soft_max, kernel_soft_max_4;
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16_q1;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bm;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bn;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
     cl_kernel kernel_set_rows_f32, kernel_set_rows_f16;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
@@ -439,6 +451,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_mul_mv_mxfp4_f32;
     cl_kernel kernel_im2col_f32, kernel_im2col_f16;
     cl_kernel kernel_argsort_f32_i32;
     cl_kernel kernel_sum_rows_f32;
@@ -455,6 +468,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_conv_2d_f16_f32;
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
+    cl_kernel kernel_mul_mv_id_mxfp4_f32;
     cl_kernel kernel_mul_mm_f32_f32_l4_lm;
     cl_kernel kernel_mul_mm_f16_f32_l4_lm;
 
@@ -577,6 +591,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_transpose_32;
     cl_kernel kernel_transpose_32_16;
     cl_kernel kernel_transpose_16;
+    cl_kernel kernel_transpose_16_4x1;
 
     cl_mem A_s_d_max;            // max scale buffer size for transpose
     cl_mem A_q_d_max;            // max weight buffer size for transpose
@@ -971,6 +986,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_mxfp4_f32, "kernel_mul_mv_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_f16_f16
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1287,6 +1318,73 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // flash_attn
+    {
+        #ifdef GGML_OPENCL_EMBED_KERNELS
+                const std::string kernel_src_f16 {
+                    #include "flash_attn_f16.cl.h"
+                };
+                const std::string kernel_src_f32 {
+                    #include "flash_attn_f32.cl.h"
+                };
+                const std::string kernel_src_f32_f16 {
+                    #include "flash_attn_f32_f16.cl.h"
+                };
+        #else
+                const std::string kernel_src_f16 = read_file("flash_attn_f16.cl");
+                const std::string kernel_src_f32 = read_file("flash_attn_f32.cl");
+                const std::string kernel_src_f32_f16 = read_file("flash_attn_f32_f16.cl");
+        #endif
+
+        if (!kernel_src_f16.empty() && !kernel_src_f32.empty() && !kernel_src_f32_f16.empty()) {
+            const struct { int dk; int dv; int bm; int bn; } fa_dims[] = {
+                { 64,  64, 64, 64}, { 80,  80, 64, 32}, { 96,  96, 64, 32},
+                {112, 112, 32, 32}, {128, 128, 32, 32}, {192, 128, 16, 16},
+                {192, 192, 16, 16}, {256, 256, 16, 16},
+            };
+
+            for (size_t i = 0; i < sizeof(fa_dims)/sizeof(fa_dims[0]); ++i) {
+                const int dk = fa_dims[i].dk;
+                const int dv = fa_dims[i].dv;
+                const int bm = fa_dims[i].bm;
+                const int bn = fa_dims[i].bn;
+                std::string OPTS = compile_opts +
+                    " -D DK=" + std::to_string(dk) +
+                    " -D DV=" + std::to_string(dv) +
+                    " -D BLOCK_M=" + std::to_string(bm) +
+                    " -D BLOCK_N=" + std::to_string(bn);
+
+                cl_program prog_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f16.c_str(), OPTS);
+                cl_kernel k_f16, k_f16_q1;
+                CL_CHECK((k_f16 = clCreateKernel(prog_f16, "flash_attn_f16", &err), err));
+                CL_CHECK((k_f16_q1 = clCreateKernel(prog_f16, "flash_attn_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f16[{dk, dv}] = k_f16;
+                backend_ctx->kernels_flash_attn_f16_q1[{dk, dv}] = k_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f16));
+
+                cl_program prog_f32 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32.c_str(), OPTS);
+                cl_kernel k_f32, k_f32_q1;
+                CL_CHECK((k_f32 = clCreateKernel(prog_f32, "flash_attn_f32", &err), err));
+                CL_CHECK((k_f32_q1 = clCreateKernel(prog_f32, "flash_attn_f32_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32[{dk, dv}] = k_f32;
+                backend_ctx->kernels_flash_attn_f32_q1[{dk, dv}] = k_f32_q1;
+                CL_CHECK(clReleaseProgram(prog_f32));
+
+                cl_program prog_f32_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32_f16.c_str(), OPTS);
+                cl_kernel k_f32_f16, k_f32_f16_q1;
+                CL_CHECK((k_f32_f16 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16", &err), err));
+                CL_CHECK((k_f32_f16_q1 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32_f16[{dk, dv}] = k_f32_f16;
+                backend_ctx->kernels_flash_attn_f32_f16_q1[{dk, dv}] = k_f32_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f32_f16));
+
+                backend_ctx->kernels_flash_attn_bm[{dk, dv}] = bm;
+                backend_ctx->kernels_flash_attn_bn[{dk, dv}] = bn;
+            }
+            GGML_LOG_CONT(".");
+        }
+    }
+
     // argsort
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1611,6 +1709,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_id_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_id_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_id_mxfp4_f32, "kernel_mul_mv_id_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -1628,6 +1742,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_transpose_32_16 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32_16", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_32    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_16    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16_4x1    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_4x1", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -2104,6 +2219,9 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &backend_ctx->max_alloc_size, NULL);
     GGML_LOG_INFO("ggml_opencl: max mem alloc size: %zu MB\n", backend_ctx->max_alloc_size/1024/1024);
 
+    clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), &backend_ctx->max_workgroup_size, NULL);
+    GGML_LOG_INFO("ggml_opencl: device max workgroup size: %lu\n", backend_ctx->max_workgroup_size);
+
     // Check SVM.
     cl_device_svm_capabilities svm_caps;
     CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_SVM_CAPABILITIES, sizeof(cl_device_svm_capabilities), &svm_caps, 0));
@@ -2419,7 +2537,8 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
 }
 
 static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    GGML_UNUSED(dev);
+    ggml_backend_opencl_device_context * dev_ctx     = (ggml_backend_opencl_device_context *)dev->context;
+    ggml_backend_opencl_context *        backend_ctx = dev_ctx->backend_ctx;
 
     switch (op->op) {
         case GGML_OP_NONE:
@@ -2481,6 +2600,13 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_SCALE:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
         case GGML_OP_ADD:
+            if (op->type == GGML_TYPE_F16) {
+                const bool src0_ok = op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32;
+                const bool src1_ok = op->src[1]->type == GGML_TYPE_F16 || op->src[1]->type == GGML_TYPE_F32;
+                if (src0_ok && src1_ok) {
+                    return true;
+                }
+            }
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_SUB:
@@ -2545,13 +2671,14 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return true;
             } else if (op->src[0]->type == GGML_TYPE_F32) {
                 return op->src[1]->type == GGML_TYPE_F32;
-            } else if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+            } else if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_MXFP4 ||
                        op->src[0]->type == GGML_TYPE_Q6_K) {
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             }
             return false;
         case GGML_OP_MUL_MAT_ID:
-            if (op->src[0]->type == GGML_TYPE_Q4_0) {
+            if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+                op->src[0]->type == GGML_TYPE_MXFP4) {
                 if (op->src[1]->type == GGML_TYPE_F32) {
                     return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
                 }
@@ -2586,10 +2713,58 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         }
         case GGML_OP_IM2COL:
             return true;
-        case GGML_OP_ARGSORT:
-            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_ARGSORT: {
+            cl_kernel kernel = backend_ctx->kernel_argsort_f32_i32;
+            int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+
+            int cols = 1;
+            while (cols < op->ne[0]) {
+                cols *= 2;
+            }
+
+            return cols <= max_workgroup_size && op->src[0]->type == GGML_TYPE_F32;
+        }
         case GGML_OP_SUM_ROWS:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                if (op->src[4]) {
+                    return false;
+                }
+
+                const ggml_tensor * q = op->src[0];
+                const ggml_tensor * k = op->src[1];
+                const ggml_tensor * v = op->src[2];
+
+                const int dk = q->ne[0];
+                const int dv = v->ne[0];
+
+                const struct { int dk; int dv; } supported_dims[] = {
+                    { 64,  64}, { 80,  80}, { 96,  96},
+                    {112, 112}, {128, 128}, {192, 128},
+                    {192, 192}, {256, 256},
+                };
+
+                bool dims_supported = false;
+                for (size_t i = 0; i < sizeof(supported_dims)/sizeof(supported_dims[0]); ++i) {
+                    if (supported_dims[i].dk == dk && supported_dims[i].dv == dv) {
+                        dims_supported = true;
+                        break;
+                    }
+                }
+                if (!dims_supported) {
+                    return false;
+                }
+
+                const bool is_f32_f32 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F32 &&
+                                        v->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+                const bool is_f16_f16 = q->type == GGML_TYPE_F16 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
+                const bool is_f32_f16 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
+                return is_f32_f32 || is_f16_f16 || is_f32_f16;
+            }
         default:
             return false;
     }
@@ -2937,7 +3112,10 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         // cl_mem qT_d = clCreateBuffer(context, CL_MEM_READ_WRITE, q_size_bytes, NULL, &err);
         CL_CHECK(err);
 
-        // size_t d_size_bytes = M * (K / 32) / 2 * sizeof(float);
+        bool K_tile_trans = true;
+        if ((K / 32) % 4 != 0){
+            K_tile_trans =false;
+        }
         size_t d_size_bytes = M * (K / 32) * 2;
         region.origin = 0;
         region.size = d_size_bytes;
@@ -2978,10 +3156,15 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
         CL_CHECK(err);
 
-        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
         memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        if (K_tile_trans) {
+            img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32 / 4;
+        } else {
+            img_fmt_1d = { CL_R, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32;
+        }
         img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-        img_desc_1d.image_width = M * K / 32 / 4;
         img_desc_1d.buffer = extra->d;
         d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
         CL_CHECK(err);
@@ -3017,6 +3200,10 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         int width_s = K / 32 / 4;
 
         kernel = backend_ctx->kernel_transpose_16;
+        if (!K_tile_trans) {
+            kernel = backend_ctx->kernel_transpose_16_4x1;
+            width_s = K / 32;
+        }
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &dT_d_image1D));
         CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_s));
@@ -3717,34 +3904,30 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
-    GGML_ASSERT(src0->type == src1->type);
-    GGML_ASSERT(src0->type == dst->type);
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
-
-    const int  ne00 = src0->ne[0];
-    const int  ne01 = src0->ne[1];
-    const int  ne02 = src0->ne[2];
-    const int  ne03 = src0->ne[3];
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
 
     const cl_ulong nb00 = src0->nb[0];
     const cl_ulong nb01 = src0->nb[1];
     const cl_ulong nb02 = src0->nb[2];
     const cl_ulong nb03 = src0->nb[3];
 
-    const int  ne10 = src1->ne[0];
-    const int  ne11 = src1->ne[1];
-    const int  ne12 = src1->ne[2];
-    const int  ne13 = src1->ne[3]; UNUSED(ne13);
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
 
     const cl_ulong nb10 = src1->nb[0];
     const cl_ulong nb11 = src1->nb[1];
     const cl_ulong nb12 = src1->nb[2];
-    const cl_ulong nb13 = src1->nb[3]; UNUSED(nb13);
+    const cl_ulong nb13 = src1->nb[3];
 
-    const int  ne0  = dst->ne[0];
-    const int  ne1  = dst->ne[1];
-    const int  ne2  = dst->ne[2];
-    const int  ne3  = dst->ne[3];
+    const int ne0  = dst->ne[0];
+    const int ne1  = dst->ne[1];
+    const int ne2  = dst->ne[2];
+    const int ne3  = dst->ne[3];
 
     const cl_ulong nb0  = dst->nb[0];
     const cl_ulong nb1  = dst->nb[1];
@@ -3761,68 +3944,114 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     cl_ulong offset1 = extra1->offset + src1->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
 
-    bool bcast_row = false;
     cl_kernel kernel;
 
-    if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-        GGML_ASSERT(ggml_is_contiguous(src0));
+    const bool bcast_row = ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0;
 
-        // src1 is a row
+    if (bcast_row) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
         GGML_ASSERT(ne11 == 1);
+    }
 
-        bcast_row = true;
-        int ne = ne00 / 4;
-
-        if (src0->type == GGML_TYPE_F32) {
+    if (dst->type == GGML_TYPE_F32) {
+        GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32);
+        if (bcast_row) {
             kernel = backend_ctx->kernel_add_row;
+            const int ne = ne00 / 4;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
         } else {
-            kernel = backend_ctx->kernel_add_row_f16;
-        }
-
-        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
-        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
-        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
-        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
-        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
-    } else {
-        if (src0->type == GGML_TYPE_F32) {
             kernel = backend_ctx->kernel_add;
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+            CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+            CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+            CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+            CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+            CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+        }
+    } else if (dst->type == GGML_TYPE_F16) {
+        GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32);
+        GGML_ASSERT(src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
+        const int type_src0 = (src0->type == GGML_TYPE_F32);
+        const int type_src1 = (src1->type == GGML_TYPE_F32);
+        if (bcast_row) {
+            kernel = backend_ctx->kernel_add_row_f16;
+            const int ne = ne00 / 4;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+            CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &type_src0));
+            CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),      &type_src1));
         } else {
             kernel = backend_ctx->kernel_add_f16;
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+            CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+            CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+            CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+            CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+            CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+            CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &type_src0));
+            CL_CHECK(clSetKernelArg(kernel, 31, sizeof(int),      &type_src1));
         }
-
-        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
-        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
-        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
-        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
-        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
-        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
-        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
-        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
-        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
-        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
-        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
-        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
-        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
-        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
-        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
-        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
-        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
-        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
-        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
-        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
-        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
-        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
-        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
-        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
-        CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
-        CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
-        CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+    } else {
+        GGML_ASSERT(false && "unsupported data types for add");
     }
 
     if (bcast_row) {
@@ -3832,13 +4061,13 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
 
         size_t * local_work_size_ptr = local_work_size;
         if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
-            local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+            local_work_size_ptr = nullptr;
         }
 
-        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+        backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size_ptr, dst);
     } else {
         unsigned int nth = MIN(64, ne0);
-        size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
         size_t local_work_size[] = {nth, 1, 1};
 
         backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
@@ -5351,6 +5580,133 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
 }
 
+static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, const ggml_tensor * k, ggml_tensor * dst) {
+    const ggml_tensor * v = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+    GGML_ASSERT(q->extra);
+    GGML_ASSERT(k->extra);
+    GGML_ASSERT(v->extra);
+    GGML_ASSERT(dst->extra);
+    if (mask) {
+        GGML_ASSERT(mask->extra);
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int n_q = q->ne[1];
+    const int n_kv = k->ne[1];
+    const int d_head_q = q->ne[0];
+    const int d_head_v = v->ne[0];
+    const int n_head = q->ne[2];
+    const int n_head_kv = k->ne[2];
+    const int n_batch = q->ne[3];
+
+    cl_kernel kernel = NULL;
+
+    const bool is_f16 = q->type == GGML_TYPE_F16;
+    const bool is_mixed = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16;
+    const std::pair<int, int> dk_dv = {d_head_q, d_head_v};
+
+    if (n_q == 1) {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16_q1.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16_q1.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32_q1.at(dk_dv);
+        }
+    } else {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32.at(dk_dv);
+        }
+    }
+    GGML_ASSERT(kernel != NULL);
+
+    ggml_tensor_extra_cl * extra_q = (ggml_tensor_extra_cl *)q->extra;
+    ggml_tensor_extra_cl * extra_k = (ggml_tensor_extra_cl *)k->extra;
+    ggml_tensor_extra_cl * extra_v = (ggml_tensor_extra_cl *)v->extra;
+    ggml_tensor_extra_cl * extra_o = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl * extra_mask = mask ? (ggml_tensor_extra_cl *)mask->extra : NULL;
+
+    cl_ulong offset_q = extra_q->offset + q->view_offs;
+    cl_ulong offset_k = extra_k->offset + k->view_offs;
+    cl_ulong offset_v = extra_v->offset + v->view_offs;
+    cl_ulong offset_o = extra_o->offset + dst->view_offs;
+    cl_mem   mask_buffer = extra_mask ? extra_mask->data_device : NULL;
+    cl_ulong offset_mask = extra_mask ? extra_mask->offset + mask->view_offs : 0;
+
+    const cl_ulong q_nb1 = q->nb[1], q_nb2 = q->nb[2], q_nb3 = q->nb[3];
+    const cl_ulong k_nb1 = k->nb[1], k_nb2 = k->nb[2], k_nb3 = k->nb[3];
+    const cl_ulong v_nb1 = v->nb[1], v_nb2 = v->nb[2], v_nb3 = v->nb[3];
+    const cl_ulong o_nb1 = dst->nb[1], o_nb2 = dst->nb[2], o_nb3 = dst->nb[3];
+    const cl_ulong mask_nb1 = mask ? mask->nb[1] : 0;
+    const cl_ulong mask_nb2 = mask ? mask->nb[2] : 0;
+    const cl_ulong mask_nb3 = mask ? mask->nb[3] : 0;
+    const int mask_ne2 = mask ? mask->ne[2] : 0;
+    const int mask_ne3 = mask ? mask->ne[3] : 0;
+
+    float scale, max_bias, logit_softcap;
+    const float * params = (const float *)dst->op_params;
+    scale         = params[0];
+    max_bias      = params[1];
+    logit_softcap = params[2];
+
+    const int is_causal = (mask == NULL && n_q > 1 && n_q == n_kv);
+
+    const int n_head_log2_val = n_head > 0 ? 1u << (int)floorf(log2f((float)n_head)) : 0;
+    const float n_head_log2_f = n_head_log2_val > 0 ? (float)n_head_log2_val : 1.0f;
+    const float m0 = powf(2.0f, -(max_bias) / n_head_log2_f);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2_f);
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra_q->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset_q));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra_k->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset_k));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extra_v->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset_v));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem),   &extra_o->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offset_o));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),      &n_q));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),     &n_kv));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),     &is_causal));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &n_head));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &q_nb1)); CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &q_nb2)); CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &q_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &k_nb1)); CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &k_nb2)); CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &k_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &v_nb1)); CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &v_nb2)); CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &v_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &o_nb1)); CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &o_nb2)); CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &o_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(int),      &n_head_log2_val));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(float),    &logit_softcap));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &n_head_kv));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_mem),   &mask_buffer));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(cl_ulong), &offset_mask));
+    CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_ulong), &mask_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 34, sizeof(cl_ulong), &mask_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 35, sizeof(cl_ulong), &mask_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 36, sizeof(int),      &mask_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 37, sizeof(int),      &mask_ne3));
+
+    if (n_q == 1) {
+        const size_t wg_size = 64;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    } else {
+        const int block_m = backend_ctx->kernels_flash_attn_bm.at(dk_dv);
+        const size_t wg_size = block_m;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { (size_t)((n_q + block_m - 1) / block_m) * wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    }
+}
+
 static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -6205,11 +6561,47 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
             CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
             break;
+        case GGML_TYPE_MXFP4: {
+            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float)*nth0,nullptr));
+            break;
+        }
         default:
             GGML_ASSERT(false && "not implemented");
     }
 
-    if (src0t == GGML_TYPE_Q4_0 ||
+    if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_MXFP4 ||
         src0t == GGML_TYPE_Q4_1 ||
         src0t == GGML_TYPE_Q8_0 ||
         src0t == GGML_TYPE_Q2_K) {
@@ -6258,10 +6650,12 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
     ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
 
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
     cl_ulong offset1 = extra1->offset + src1->view_offs;
     cl_ulong offset2 = extra2->offset + src2->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
@@ -6276,7 +6670,9 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
     const int ne03 = src0->ne[3];
 
     const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
     const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
 
     const int ne10 = src1->ne[0];
     const int ne11 = src1->ne[1];
@@ -6285,6 +6681,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
     const cl_ulong nb11 = src1->nb[1];
     const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
 
     const int ne20 = src2->ne[0];
     const int ne21 = src2->ne[1];
@@ -6352,6 +6749,49 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
             break;
         }
+        case GGML_TYPE_MXFP4: {
+            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs,nullptr));
+
+            break;
+        }
         default:
             GGML_ASSERT(false && "not implemented");;
     }
@@ -7423,6 +7863,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_sum_rows;
             break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_flash_attn(backend, tensor->src[0], tensor->src[1], tensor);
+            return true;
         default:
             return false;
     }
diff --git a/ggml/src/ggml-opencl/kernels/add.cl b/ggml/src/ggml-opencl/kernels/add.cl
index 8bc926c88931f..509bf17344ea6 100644
--- a/ggml/src/ggml-opencl/kernels/add.cl
+++ b/ggml/src/ggml-opencl/kernels/add.cl
@@ -112,7 +112,9 @@ kernel void kernel_add_f16(
         ulong nb0,
         ulong nb1,
         ulong nb2,
-        ulong nb3
+        ulong nb3,
+        int type_src0,
+        int type_src1
 ) {
     src0 = src0 + offset0;
     src1 = src1 + offset1;
@@ -132,25 +134,57 @@ kernel void kernel_add_f16(
 
     for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
         const int i10 = i0 % ne10;
-        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) + *((global half *)(src1_ptr + i10*nb10));
+
+        half v0, v1;
+        if (type_src0 == 1) {
+            v0 = convert_half(*((global float *)(src0_ptr + i0*nb00)));
+        } else {
+            v0 = *((global half *)(src0_ptr + i0*nb00));
+        }
+
+        if (type_src1 == 1) {
+            v1 = convert_half(*((global float *)(src1_ptr + i10*nb10)));
+        } else {
+            v1 = *((global half *)(src1_ptr + i10*nb10));
+        }
+
+        *((global half *)(dst_ptr + i0*nb0)) = v0 + v1;
     }
 }
 
 kernel void kernel_add_row_f16(
-        global half4 * src0,
+        global char * src0,
         ulong  offset0,
-        global half4 * src1,
+        global char * src1,
         ulong  offset1,
         global half4 * dst,
         ulong  offsetd,
-        int ne
+        int ne,
+        int type_src0,
+        int type_src1
 ) {
-    src0 = (global half4*)((global char*)src0 + offset0);
-    src1 = (global half4*)((global char*)src1 + offset1);
     dst = (global half4*)((global char*)dst + offsetd);
 
     // This performs better than using %.
     uint gid = get_global_id(0);
     uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
-    dst[gid] = src0[gid] + src1[idx1];
+
+    half4 v0, v1;
+    if (type_src0 == 1) {
+        global float4* src0_f32 = (global float4*)((global char*)src0 + offset0);
+        v0 = convert_half4(src0_f32[gid]);
+    } else {
+        global half4* src0_f16 = (global half4*)((global char*)src0 + offset0);
+        v0 = src0_f16[gid];
+    }
+
+    if (type_src1 == 1) {
+        global float4* src1_f32 = (global float4*)((global char*)src1 + offset1);
+        v1 = convert_half4(src1_f32[idx1]);
+    } else {
+        global half4* src1_f16 = (global half4*)((global char*)src1 + offset1);
+        v1 = src1_f16[idx1];
+    }
+
+    dst[gid] = v0 + v1;
 }
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl
new file mode 100644
index 0000000000000..fea06867e1020
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl
@@ -0,0 +1,343 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define DATA_TYPE half
+#define DATA_TYPE4 half4
+#define CONVERT_ACC4(x) convert_float4(x)
+#define CONVERT_DATA4(x) convert_half4(x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f16(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f16_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl
new file mode 100644
index 0000000000000..2d657327d6460
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl
@@ -0,0 +1,343 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define DATA_TYPE float
+#define DATA_TYPE4 float4
+#define CONVERT_ACC4(x) (x)
+#define CONVERT_DATA4(x) (x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f32(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f32_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl b/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
new file mode 100644
index 0000000000000..7067bd2591fa7
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
@@ -0,0 +1,346 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define Q_DATA_TYPE4 float4
+#define KV_DATA_TYPE4 half4
+#define O_DATA_TYPE4 float4
+#define MASK_DATA_TYPE half
+#define CONVERT_Q_ACC4(x) (x)
+#define CONVERT_KV_ACC4(x) convert_float4(x)
+#define CONVERT_O_DATA4(x) (x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f32_f16(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local KV_DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local KV_DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global KV_DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_KV_ACC4(l_v[j][i]) + p1 * CONVERT_KV_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_O_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (O_DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f32_f16_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
+        const global KV_DATA_TYPE4* v_ptr = (const global KV_DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_KV_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_O_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (O_DATA_TYPE4)(0.0f);
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl b/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl
new file mode 100644
index 0000000000000..d50bd1fc4285d
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl
@@ -0,0 +1,189 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint) x << 23;
+    }
+
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_mv_mxfp4_f32(
+    global char * src0,
+    global char * src1,
+    global char * dst,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3,
+    local  char * shmem
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11 = idx % ne11;
+    int i12 = iid1;
+
+    int i1 = idx;
+    int i2 = i12;
+
+    global char * src0_cur = src0 + i02*nb02;
+    global char * src1_cur = src1 + i11*nb11 + i12*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    mul_mv_mxfp4_f32(src0_cur, src1_cur, dst_cur,
+        ne00, nb01, nb02, nb03, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shmem);
+}
diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl b/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl
new file mode 100644
index 0000000000000..9a4d4b9bad1dd
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl
@@ -0,0 +1,144 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint) x << 23;
+    }
+
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/transpose.cl b/ggml/src/ggml-opencl/kernels/transpose.cl
index a11490b304c5b..536dd560a917b 100644
--- a/ggml/src/ggml-opencl/kernels/transpose.cl
+++ b/ggml/src/ggml-opencl/kernels/transpose.cl
@@ -24,6 +24,26 @@ kernel void kernel_transpose_16(
     write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
 }
 
+// Padded kernel for irregular shape
+kernel void kernel_transpose_16_4x1(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int j_2 = j << 2;
+
+    half temp0 = read_imageh(input, (j_2 + 0) * cols + i).x;
+    half temp1 = read_imageh(input, (j_2 + 1) * cols + i).x;
+    half temp2 = read_imageh(input, (j_2 + 2) * cols + i).x;
+    half temp3 = read_imageh(input, (j_2 + 3) * cols + i).x;
+
+    write_imageh(output, i * rows + j, (half4)(temp0, temp1, temp2, temp3));
+}
+
 // 32-bit transpose, loading/storing a 4x4 tile of elements
 kernel void kernel_transpose_32(
     __read_only image1d_buffer_t input,
diff --git a/ggml/src/ggml-opt.cpp b/ggml/src/ggml-opt.cpp
index a3c82d6757714..e078ad14a39c4 100644
--- a/ggml/src/ggml-opt.cpp
+++ b/ggml/src/ggml-opt.cpp
@@ -64,9 +64,11 @@ struct ggml_opt_context {
     int32_t opt_i              = 0;
     bool    loss_per_datapoint = false;
 
-    ggml_opt_get_optimizer_params get_opt_pars = nullptr;
-    void * get_opt_pars_ud                     = nullptr;
-    struct ggml_tensor * adamw_params          = nullptr;
+    ggml_opt_get_optimizer_params get_opt_pars    = nullptr;
+    void *                        get_opt_pars_ud = nullptr;
+    struct ggml_tensor *          opt_step_params = nullptr; // Stores output of get_opt_pars.
+
+    enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
 };
 
 struct ggml_opt_result {
@@ -229,9 +231,13 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
     result.adamw.eps   = 1e-8f;
     result.adamw.wd    = 0.0f;
 
+    result.sgd.alpha   = 1e-3f;
+    result.sgd.wd      = 0.0f;
+
     return result;
 }
 
+
 struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata) {
     return *((struct ggml_opt_optimizer_params *) userdata);
 }
@@ -249,6 +255,7 @@ struct ggml_opt_params ggml_opt_default_params(
         /*opt_period      =*/ 1,
         /*get_opt_pars    =*/ ggml_opt_get_default_optimizer_params,
         /*get_opt_pars_ud =*/ nullptr,
+        /*optimizer       =*/ GGML_OPT_OPTIMIZER_TYPE_ADAMW,
     };
 }
 
@@ -316,9 +323,14 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
     GGML_ASSERT(opt_ctx->ctx_compute && "no compute context set, either use static graphs or set one with ggml_opt_prepare_alloc");
     GGML_ASSERT((!opt_ctx->static_graphs || opt_ctx->inputs->data) && "when using static graphs the inputs must be allocated statically");
 
+    const enum ggml_opt_optimizer_type optimizer = opt_ctx->optimizer;
+
     const bool accumulate = opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD &&
         !(opt_ctx->static_graphs && opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period == 1);
 
+    const bool need_momenta = opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT &&
+        opt_ctx->optimizer == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+
     ggml_set_input(opt_ctx->inputs);
     ggml_set_output(opt_ctx->outputs);
 
@@ -340,8 +352,7 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
         //   - pred (if using static graphs)
         //   - ncorrect (if using static graphs, 2 tensors).
         constexpr size_t n_loss = 1;
-        const size_t tensors_per_param = (accumulate ? 1 : 0) +
-            (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t tensors_per_param = (accumulate ? 1 : 0) + (need_momenta ? 2 : 0);
         const size_t tensors_const = opt_ctx->static_graphs ? 9 : 0;
         const size_t size_meta = (n_loss + tensors_per_param*n_param + tensors_const) * ggml_tensor_overhead();
         struct ggml_init_params params = {
@@ -458,7 +469,7 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
             }
         }
 
-        if (opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
+        if (need_momenta && opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
             opt_ctx->grad_m.resize(n_nodes);
             opt_ctx->grad_v.resize(n_nodes);
             for (int i = 0; i < n_nodes; ++i) {
@@ -492,23 +503,36 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
     // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
     opt_ctx->gb_opt = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gb_grad, /*force_grads =*/ true);
 
-    opt_ctx->adamw_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, 7);
-    ggml_set_input(opt_ctx->adamw_params);
-    ggml_set_name(opt_ctx->adamw_params, "adamw_params");
-
+    opt_ctx->opt_step_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, need_momenta ? 7 : 2);
+    ggml_tensor * adamw_params = opt_ctx->opt_step_params;
+    ggml_set_input(adamw_params);
+    const char * optimizer_name = ggml_opt_optimizer_name(opt_ctx->optimizer);
+    ggml_format_name(adamw_params, "%s_params", optimizer_name);
     for (int i = opt_ctx->gf->n_nodes-1; i >= 0; --i) {
         struct ggml_tensor * node = opt_ctx->gb_opt->nodes[i];
         struct ggml_tensor * grad = ggml_graph_get_grad(opt_ctx->gb_opt, node);
 
         if (grad && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
-            struct ggml_tensor * m        = opt_ctx->grad_m[i];
-            struct ggml_tensor * v        = opt_ctx->grad_v[i];
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, opt_ctx->adamw_params);
-
-            ggml_set_name(m,        (std::string("AdamW m for ")    + std::string(node->name)).c_str());
-            ggml_set_name(v,        (std::string("AdamW v for ")    + std::string(node->name)).c_str());
-            ggml_set_name(opt_step, (std::string("AdamW step for ") + std::string(node->name)).c_str());
-
+            struct ggml_tensor * m = nullptr;
+            struct ggml_tensor * v = nullptr;
+            if (need_momenta) {
+                m = opt_ctx->grad_m[i];
+                v = opt_ctx->grad_v[i];
+                ggml_format_name(m, "AdamW m for %s", node->name);
+                ggml_format_name(v, "AdamW v for %s", node->name);
+            }
+            struct ggml_tensor * opt_step;
+            switch (optimizer) {
+                case GGML_OPT_OPTIMIZER_TYPE_ADAMW:
+                    opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, adamw_params);
+                    break;
+                case GGML_OPT_OPTIMIZER_TYPE_SGD:
+                    opt_step = ggml_opt_step_sgd(opt_ctx->ctx_compute, node, grad, adamw_params);
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+            ggml_format_name(opt_step, "%s step for %s", optimizer_name, node->name);
             ggml_build_forward_expand(opt_ctx->gb_opt, opt_step);
         }
     }
@@ -534,6 +558,7 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     result->opt_period       = params.opt_period;
     result->get_opt_pars     = params.get_opt_pars;
     result->get_opt_pars_ud  = params.get_opt_pars_ud;
+    result->optimizer        = params.optimizer;
 
     GGML_ASSERT(result->opt_period >= 1);
 
@@ -756,29 +781,43 @@ void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward) {
 void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result) {
     GGML_ASSERT(opt_ctx->eval_ready);
     if (opt_ctx->allocated_graph == opt_ctx->gb_opt) {
-        struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
-
-        GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
-        GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
-        GGML_ASSERT(opt_pars.adamw.eps   >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.wd    >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.wd    <= 1.0f);
-
-        // beta1, beta2 after applying warmup
-        const float beta1h = 1.0f/(1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
-        const float beta2h = 1.0f/(1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
-
-        float * adamw_par_data = ggml_get_data_f32(opt_ctx->adamw_params);
-        adamw_par_data[0] = opt_pars.adamw.alpha;
-        adamw_par_data[1] = opt_pars.adamw.beta1;
-        adamw_par_data[2] = opt_pars.adamw.beta2;
-        adamw_par_data[3] = opt_pars.adamw.eps;
-        adamw_par_data[4] = opt_pars.adamw.wd;
-        adamw_par_data[5] = beta1h;
-        adamw_par_data[6] = beta2h;
+        const ggml_opt_optimizer_params & opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
+
+        switch (opt_ctx->optimizer) {
+            case GGML_OPT_OPTIMIZER_TYPE_ADAMW: {
+                GGML_ASSERT(opt_pars.adamw.alpha > 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
+                GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
+                GGML_ASSERT(opt_pars.adamw.eps >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.wd >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.wd <= 1.0f);
+
+                // beta1, beta2 after applying warmup
+                const float beta1h = 1.0f / (1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
+                const float beta2h = 1.0f / (1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
+
+                float * adamw_par_data = ggml_get_data_f32(opt_ctx->opt_step_params);
+                adamw_par_data[0] = opt_pars.adamw.alpha;
+                adamw_par_data[1] = opt_pars.adamw.beta1;
+                adamw_par_data[2] = opt_pars.adamw.beta2;
+                adamw_par_data[3] = opt_pars.adamw.eps;
+                adamw_par_data[4] = opt_pars.adamw.wd;
+                adamw_par_data[5] = beta1h;
+                adamw_par_data[6] = beta2h;
+            } break;
+            case GGML_OPT_OPTIMIZER_TYPE_SGD: {
+                GGML_ASSERT(opt_pars.sgd.alpha > 0.0f);
+                GGML_ASSERT(opt_pars.sgd.wd >= 0.0f);
+                GGML_ASSERT(opt_pars.sgd.wd <= 1.0f);
+                float * sgd = ggml_get_data_f32(opt_ctx->opt_step_params);
+                sgd[0] = opt_pars.sgd.alpha;
+                sgd[1] = opt_pars.sgd.wd;
+            } break;
+            default:
+                GGML_ABORT("fatal error");
+        }
     }
 
     ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
@@ -963,6 +1002,7 @@ void ggml_opt_fit(
         ggml_tensor                   * outputs,
         ggml_opt_dataset_t              dataset,
         enum ggml_opt_loss_type         loss_type,
+        enum ggml_opt_optimizer_type    optimizer,
         ggml_opt_get_optimizer_params   get_opt_pars,
         int64_t                         nepoch,
         int64_t                         nbatch_logical,
@@ -993,6 +1033,7 @@ void ggml_opt_fit(
     params.opt_period      = opt_period;
     params.get_opt_pars    = get_opt_pars;
     params.get_opt_pars_ud = &epoch;
+    params.optimizer       = optimizer;
     ggml_opt_context_t opt_ctx = ggml_opt_init(params);
 
     // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
@@ -1035,3 +1076,18 @@ void ggml_opt_fit(
     ggml_opt_result_free(result_train);
     ggml_opt_result_free(result_val);
 }
+
+enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t c) {
+    return c->optimizer;
+}
+
+GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type o) {
+    switch (o) {
+        case GGML_OPT_OPTIMIZER_TYPE_ADAMW:
+            return "adamw";
+        case GGML_OPT_OPTIMIZER_TYPE_SGD:
+            return "sgd";
+        default:
+            return "undefined";
+    };
+}
diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 94f6405ca1e05..727932123e41b 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -566,7 +566,7 @@ static float make_q3_quants(int n, int nmax, const float * GGML_RESTRICT x, int8
         for (int i = 0; i < n; ++i) {
             L[i] += nmax;
         }
-        return sumlx / suml2;
+        return suml2 > 0.0f ? sumlx / suml2 : 0.0f;
     }
     for (int i = 0; i < n; ++i) {
         int l = nearest_int(iscale * x[i]);
@@ -901,7 +901,7 @@ static float make_qp_quants(int n, int nmax, const float * GGML_RESTRICT x, uint
     for (int i = 0; i < n; ++i) {
         max = MAX(max, x[i]);
     }
-    if (!max) { // all zero
+    if (max < GROUP_MAX_EPS) { // all zero
         for (int i = 0; i < n; ++i) { L[i] = 0; }
         return 0.f;
     }
@@ -966,7 +966,7 @@ static float make_qp_quants(int n, int nmax, const float * GGML_RESTRICT x, uint
             break;
         }
     }
-    return sumlx/suml2;
+    return suml2 > 0.0f ? sumlx / suml2 : 0.0f;
 }
 
 static void quantize_row_q2_K_impl(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int k, const float * GGML_RESTRICT quant_weights) {
@@ -4266,7 +4266,7 @@ static void quantize_row_iq1_s_impl(const float * GGML_RESTRICT x, void * GGML_R
                     sumw[j+1] = sumw[j] + weight[i];
                 }
             }
-            float best_score = -FLT_MIN, scale = max;
+            float best_score = -FLT_MAX, scale = max;
             int besti1 = -1, besti2 = -1, best_shift = 0;
             for (int i1 = 0; i1 <= block_size; ++i1) {
                 for (int i2 = i1; i2 <= block_size; ++i2) {
@@ -4442,7 +4442,7 @@ static void quantize_row_iq1_m_impl(const float * GGML_RESTRICT x, void * GGML_R
                 idx[2*j] = j;
             }
             qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
-            float best_score = -FLT_MIN, scale = max;
+            float best_score = -FLT_MAX, scale = max;
             int besti1 = -1, besti2 = -1, best_k = -1;
             // 0: +, +
             // 1: +, -
diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
index df6ba54076d20..e84ff93efc32c 100644
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -29,9 +29,12 @@
 #include <cstring>
 #include <fstream>
 #include <filesystem>
+#include <algorithm>
 
 namespace fs = std::filesystem;
 
+static constexpr size_t MAX_CHUNK_SIZE = 1024ull * 1024ull * 1024ull; // 1 GiB
+
 #ifdef _WIN32
 typedef SOCKET sockfd_t;
 using ssize_t = __int64;
@@ -323,11 +326,14 @@ static std::shared_ptr<socket_t> create_server_socket(const char * host, int por
 static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
     size_t bytes_sent = 0;
     while (bytes_sent < size) {
-        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size - bytes_sent, 0);
+        size_t size_to_send = std::min(size - bytes_sent, MAX_CHUNK_SIZE);
+        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size_to_send, 0);
         if (n < 0) {
+            GGML_LOG_ERROR("send failed (bytes_sent=%zu, size_to_send=%zu)\n",
+                           bytes_sent, size_to_send);
             return false;
         }
-        bytes_sent += n;
+        bytes_sent += (size_t)n;
     }
     return true;
 }
@@ -335,11 +341,18 @@ static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
 static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
     size_t bytes_recv = 0;
     while (bytes_recv < size) {
-        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size - bytes_recv, 0);
-        if (n <= 0) {
+        size_t size_to_recv = std::min(size - bytes_recv, MAX_CHUNK_SIZE);
+        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size_to_recv, 0);
+        if (n < 0) {
+            GGML_LOG_ERROR("recv failed (bytes_recv=%zu, size_to_recv=%zu)\n",
+                           bytes_recv, size_to_recv);
+            return false;
+        }
+        if (n == 0) {
+            GGML_LOG_ERROR("recv returned 0 (peer closed?)\n");
             return false;
         }
-        bytes_recv += n;
+        bytes_recv += (size_t)n;
     }
     return true;
 }
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 3992dad01d5c1..12dd5dd2e6287 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -2705,9 +2705,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                                                 " : converting src1 to fp16");
 
         // iterate tensor dims and find the slowest moving dim and stride
-        int64_t last_dim=0;
-        int64_t last_str=0;
-        int64_t largest_str=0;
+        int last_dim=0;
+        int last_str=0;
+        size_t largest_str=0;
         for(int i = 0; i< 4; i++){
             // last stride is always the largest
             if(src1->nb[i] == largest_str){
@@ -2783,7 +2783,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             auto launch_gemm_for_batches = [&ctx, queue](const sycl::half *src0,
                                                 const sycl::half *src1, float *dst,
                                                 int64_t a0, int64_t a1, int64_t batcha,
-                                                int64_t b0, int64_t b1, int64_t batchb,
+                                                int64_t /*b0*/, int64_t b1, int64_t batchb,
                                                 int64_t sa0, int64_t sa1, int64_t sa2,
                                                 int64_t sb0, int64_t sb1, int64_t sb2,
                                                 int64_t sd2) {
@@ -2832,14 +2832,26 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                 }
             };
 
-            bool cont_batches_a = nb02 * ne02 == nb03;
-            bool cont_batches_b = nb12 * ne12 == nb13;
-            if (cont_batches_a && cont_batches_b) {
+            const bool cont_batches_dim2_a = nb02 * ne02 == nb03;
+            const bool cont_batches_dim2_b = nb12 * ne12 == nb13;
+            const bool cont_batches_dim3_a = ne02 == 1 && nb02 * ne01 == nb03;
+            const bool cont_batches_dim3_b = ne12 == 1 && nb12 * ne11 == nb13;
+            if (cont_batches_dim2_a && cont_batches_dim2_b) {
+                // A batch is considered contiguous if the dimension 2 is not strided
                 int64_t batches0 = ne02 * ne03;
                 int64_t batches1 = ne12 * ne13;
                 launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
                         ne10, ne11, batches1, str_a0, str_a1, str_a2, str_b0, str_b1,
                         str_b2, nb2 / sizeof(float));
+            } else if (cont_batches_dim3_a && cont_batches_dim3_b) {
+                // This case is similar to the one above with the difference that only the batch in dimension 3 is used and the dimension 2 is of size 1.
+                int64_t batches0 = ne02 * ne03;
+                int64_t batches1 = ne12 * ne13;
+                int64_t str_a3 = nb03 / type_size_src0;
+                int64_t str_b3 = nb13 / type_size_src1;
+                launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
+                        ne10, ne11, batches1, str_a0, str_a1, str_a3, str_b0, str_b1,
+                        str_b3, nb2 / sizeof(float));
             } else {
                 for (int64_t b_a = 0; b_a < ne03; b_a++) {
                     const sycl::half *src0_f16_shifted
@@ -4215,6 +4227,15 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     // FIXME: keep a list of supported types to avoid breaking the backend when a new type is added
                     return false;
                 }
+                // TODO: The configuration below needs more work to be supported with oneDNN
+                if (ggml_is_permuted(a) && !ggml_is_contiguous(a) && a->ne[2] > 1 && a->ne[3] > 1) {
+                    return false;
+                }
+                // TODO: This specific configuration can fail with oneDNN and needs more debugging
+                if (!ggml_is_permuted(a) && ggml_is_permuted(b) && b->ne[2] > 1 && b->ne[3] > 1 &&
+                    a->ne[0] > 128 && a->ne[2] == 1 && src0_type == GGML_TYPE_F16) {
+                    return false;
+                }
                 return true;
             }
         case GGML_OP_OUT_PROD:
@@ -4370,10 +4391,11 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return true;
         case GGML_OP_UPSCALE:
             return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
-        case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_POOL_2D:
         case GGML_OP_ACC:
         case GGML_OP_PAD:
         case GGML_OP_LEAKY_RELU:
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 4070e248baa28..a5406f761274d 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -102,7 +102,9 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 
 struct ggml_backend_vk_context;
 
-#define MAX_PARAMETER_COUNT 8
+#define MAX_PARAMETER_COUNT 12
+// Max number of adds that can be fused without exceeding MAX_PARAMETER_COUNT.
+#define MAX_FUSED_ADDS (MAX_PARAMETER_COUNT - 3)
 
 struct vk_pipeline_struct {
     std::string name;
@@ -113,6 +115,8 @@ struct vk_pipeline_struct {
     uint32_t parameter_count;
     std::array<uint32_t, 3> wg_denoms;
     uint32_t align;
+    // true if fields have been set by ggml_vk_create_pipeline
+    bool initialized {};
     // set to true to request the pipeline is compiled after the dryrun
     bool needed {};
     // set to true when the shader has been compiled
@@ -225,21 +229,6 @@ enum vk_device_architecture {
     NVIDIA_PRE_TURING,
 };
 
-// HSK x HSV
-enum FaHeadSizes {
-    FA_HEAD_SIZE_64,
-    FA_HEAD_SIZE_80,
-    FA_HEAD_SIZE_96,
-    FA_HEAD_SIZE_112,
-    FA_HEAD_SIZE_128,
-    FA_HEAD_SIZE_192,
-    FA_HEAD_SIZE_192_128,
-    FA_HEAD_SIZE_256,
-    FA_HEAD_SIZE_576_512,
-    FA_HEAD_SIZE_UNSUPPORTED,
-    FA_HEAD_SIZE_COUNT = FA_HEAD_SIZE_UNSUPPORTED,
-};
-
 static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
     vk::PhysicalDeviceProperties props = device.getProperties();
 
@@ -343,6 +332,37 @@ enum vk_conv_shapes {
     CONV_SHAPE_COUNT,
 };
 
+enum dmmv_wg_sizes {
+    DMMV_WG_SIZE_SUBGROUP,
+    DMMV_WG_SIZE_LARGE,
+    DMMV_WG_SIZE_COUNT,
+};
+
+enum FaCodePath {
+    FA_SCALAR,
+    FA_COOPMAT1,
+    FA_COOPMAT2,
+};
+
+struct vk_fa_pipeline_state {
+    vk_fa_pipeline_state(uint32_t HSK, uint32_t HSV, bool small_rows, FaCodePath path, bool aligned, bool f32acc)
+        : HSK(HSK), HSV(HSV), small_rows(small_rows), path(path), aligned(aligned), f32acc(f32acc) {}
+
+    uint32_t HSK, HSV;
+    bool small_rows;
+    FaCodePath path;
+    bool aligned;
+    bool f32acc;
+
+    bool operator<(const vk_fa_pipeline_state &b) const {
+        return std::tie(HSK, HSV, small_rows, path, aligned, f32acc) <
+               std::tie(b.HSK, b.HSV, b.small_rows, b.path, b.aligned, b.f32acc);
+    }
+};
+
+static constexpr uint32_t num_argsort_pipelines = 11;
+static constexpr uint32_t max_argsort_cols = 1 << (num_argsort_pipelines-1);
+
 struct vk_device_struct {
     std::recursive_mutex mutex;
 
@@ -368,6 +388,10 @@ struct vk_device_struct {
     bool float_controls_rte_fp16;
     bool subgroup_add;
     bool subgroup_shuffle;
+    bool multi_add;
+
+    bool add_rms_fusion;
+    uint32_t partials_binding_alignment;
 
     bool integer_dot_product;
 
@@ -429,8 +453,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_quantize_q8_1;
 
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio];
@@ -448,6 +472,12 @@ struct vk_device_struct {
     vk_pipeline pipeline_mul_norepeat[2][2][2];
     vk_pipeline pipeline_div[2][2][2];
     vk_pipeline pipeline_div_norepeat[2][2][2];
+    vk_pipeline pipeline_add_rms[2][2][2];
+    vk_pipeline pipeline_add_rms_norepeat[2][2][2];
+
+    // indexed by num_additional_fused_ops == num_adds - 1
+    vk_pipeline pipeline_multi_add[MAX_FUSED_ADDS];
+    vk_pipeline pipeline_multi_add_rms[MAX_FUSED_ADDS];
 
     vk_pipeline pipeline_add_id_f32;
 
@@ -455,6 +485,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bilinear_ac_f32;
     vk_pipeline pipeline_scale_f32;
     vk_pipeline pipeline_sqr_f32;
+    vk_pipeline pipeline_sqrt_f32;
     vk_pipeline pipeline_sin_f32;
     vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_clamp_f32;
@@ -470,10 +501,13 @@ struct vk_device_struct {
     vk_pipeline pipeline_group_norm_f32;
     vk_pipeline pipeline_rms_norm_f32;
     vk_pipeline pipeline_rms_norm_mul_f32;
+    vk_pipeline pipeline_rms_norm_partials_f32;
+    vk_pipeline pipeline_rms_norm_mul_partials_f32;
     vk_pipeline pipeline_rms_norm_back_f32;
     vk_pipeline pipeline_l2_norm_f32;
 
     // [src/dst 0=fp32,1=fp16]
+    vk_pipeline pipeline_exp[2];
     vk_pipeline pipeline_gelu[2];
     vk_pipeline pipeline_gelu_erf[2];
     vk_pipeline pipeline_gelu_quick[2];
@@ -499,7 +533,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
     vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16;
-    vk_pipeline pipeline_argsort_f32;
+    vk_pipeline pipeline_argsort_f32[num_argsort_pipelines];
     vk_pipeline pipeline_sum_rows_f32;
     vk_pipeline pipeline_argmax_f32;
     vk_pipeline pipeline_count_equal_i32;
@@ -510,21 +544,17 @@ struct vk_device_struct {
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
+    vk_pipeline pipeline_opt_step_sgd_f32;
     vk_pipeline pipeline_conv2d_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
-    vk_pipeline pipeline_conv2d_dw_whcn_f32;
-    vk_pipeline pipeline_conv2d_dw_cwhn_f32;
-
-    // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
-    vk_pipeline pipeline_flash_attn_f32_f16_cm2[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
-
-    vk_pipeline pipeline_flash_attn_f32_f16_cm1[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
+    vk_pipeline pipeline_conv2d_dw_whcn_f32, pipeline_conv2d_dw_whcn_f16_f32;
+    vk_pipeline pipeline_conv2d_dw_cwhn_f32, pipeline_conv2d_dw_cwhn_f16_f32;
 
-    vk_pipeline pipeline_flash_attn_f32_f16[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
+    std::map<vk_fa_pipeline_state, vk_pipeline> pipeline_flash_attn_f32_f16[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_flash_attn_split_k_reduce;
 
-    std::unordered_map<std::string, vk_pipeline_ref> pipelines;
+    std::vector<vk_pipeline_ref> all_pipelines;
 
     std::vector<std::tuple<void*, size_t, vk_buffer>> pinned_memory;
 
@@ -555,15 +585,15 @@ struct vk_device_struct {
         compute_queue.cmd_pool.destroy(device);
         transfer_queue.cmd_pool.destroy(device);
 
-        for (auto& pipeline : pipelines) {
-            if (pipeline.second.expired()) {
+        for (auto& pipeline : all_pipelines) {
+            if (pipeline.expired()) {
                 continue;
             }
 
-            vk_pipeline pl = pipeline.second.lock();
+            vk_pipeline pl = pipeline.lock();
             ggml_vk_destroy_pipeline(device, pl);
         }
-        pipelines.clear();
+        all_pipelines.clear();
 
         device.destroyDescriptorSetLayout(dsl);
 
@@ -800,6 +830,19 @@ struct vk_op_binary_push_constants {
     float param1; float param2; int32_t param3;
 };
 
+struct vk_op_multi_add_push_constants {
+    // shape for dst
+    uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23;
+
+    // strides for srcs+dst
+    uint32_t nb[MAX_PARAMETER_COUNT][4];
+
+    uint32_t rms_partials;
+};
+// update multi_add.comp if this changes
+static_assert(MAX_PARAMETER_COUNT == 12);
+static_assert(sizeof(vk_op_multi_add_push_constants) <= 256);
+
 struct vk_op_add_id_push_constants {
     uint32_t ne0;
     uint32_t ne1;
@@ -855,7 +898,6 @@ struct vk_op_soft_max_push_constants {
 
 struct vk_op_argsort_push_constants {
     uint32_t ncols;
-    uint32_t ncols_pad;
     int32_t order;
 };
 
@@ -990,6 +1032,39 @@ struct vk_op_upscale_push_constants {
     float sf0; float sf1; float sf2; float sf3;
 };
 
+struct vk_op_sum_rows_push_constants
+{
+    uint32_t n_cols;
+    uint32_t ne01, ne02;
+    uint32_t nb01, nb02, nb03;
+    uint32_t nb11, nb12, nb13;
+    float weight;
+    uint32_t misalign_offsets;
+    uint32_t ne0_12mp, ne0_12L;
+    uint32_t ne0_1mp, ne0_1L;
+};
+
+vk_op_sum_rows_push_constants vk_op_sum_rows_push_constants_init(const ggml_tensor * src, const ggml_tensor * dst, int64_t n_cols) {
+    uint32_t type_size = (uint32_t)ggml_type_size(src->type);
+    vk_op_sum_rows_push_constants p = {};
+    p.n_cols = (uint32_t)n_cols;
+    p.ne01 = (uint32_t)src->ne[1];
+    p.ne02 = (uint32_t)src->ne[2];
+    p.nb01 = (uint32_t)src->nb[1] / type_size;
+    p.nb02 = (uint32_t)src->nb[2] / type_size;
+    p.nb03 = (uint32_t)src->nb[3] / type_size;
+    p.nb11 = (uint32_t)dst->nb[1] / type_size;
+    p.nb12 = (uint32_t)dst->nb[2] / type_size;
+    p.nb13 = (uint32_t)dst->nb[3] / type_size;
+    p.weight = 1.0f;
+    return p;
+}
+
+template <> void init_pushconst_fastdiv(vk_op_sum_rows_push_constants &p) {
+    init_fastdiv_values(p.ne01*p.ne02, p.ne0_12mp, p.ne0_12L);
+    init_fastdiv_values(p.ne01,        p.ne0_1mp,  p.ne0_1L);
+}
+
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -1110,17 +1185,23 @@ class vk_perf_logger {
             return;
         }
         if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
-            const uint64_t m    = node->src[0]->ne[1];
-            const uint64_t n    = node->src[1]->ne[1];
-            const uint64_t k    = node->src[1]->ne[0];
-            std::string    name = ggml_op_name(node->op);
-            if (n == 1) {
-                name += "_VEC m=" + std::to_string(m) + " k=" + std::to_string(k);
-            } else {
-                name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
+            const uint64_t m     = node->src[0]->ne[1];
+            const uint64_t n     = node->ne[1];
+            const uint64_t k     = node->src[1]->ne[0];
+            const uint64_t batch = node->src[1]->ne[2] * node->src[1]->ne[3];
+            std::string    name  = ggml_op_name(node->op);
+            if ((node->op == GGML_OP_MUL_MAT && n <= mul_mat_vec_max_cols) ||
+                (node->op == GGML_OP_MUL_MAT_ID && node->src[2]->ne[1] == 1)) {
+                name += "_VEC";
+            }
+            name += " ";
+            name += ggml_type_name(node->src[0]->type);
+            name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
+            if (batch > 1) {
+                name += " batch=" + std::to_string(batch);
             }
             timings[name].push_back(time);
-            flops[name].push_back(m * n * (k + (k - 1)));
+            flops[name].push_back(m * n * (k + (k - 1)) * batch);
             return;
         }
         if (node->op == GGML_OP_CONV_2D) {
@@ -1144,6 +1225,12 @@ class vk_perf_logger {
             timings[name].push_back(time);
             return;
         }
+        if (node->op == GGML_OP_RMS_NORM) {
+            std::string   name    = ggml_op_name(node->op);
+            name += "(" + std::to_string(node->ne[0]) + "," + std::to_string(node->ne[1]) + "," + std::to_string(node->ne[2]) + "," + std::to_string(node->ne[3]) + ")";
+            timings[name].push_back(time);
+            return;
+        }
         timings[ggml_op_name(node->op)].push_back(time);
     }
   private:
@@ -1158,10 +1245,25 @@ struct ggml_backend_vk_context {
 
     size_t semaphore_idx, event_idx;
     ggml_vk_garbage_collector gc;
-    size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k;
-    vk_buffer prealloc_x, prealloc_y, prealloc_split_k;
+    size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k, prealloc_size_add_rms_partials, prealloc_size_add_rms_partials_offset;
+    vk_buffer prealloc_x, prealloc_y, prealloc_split_k, prealloc_add_rms_partials;
     vk::Fence fence, almost_ready_fence;
     bool almost_ready_fence_pending {};
+    // Set before op_add and unset after op_rms_norm to indicate that the add should
+    // write partial sums to accumulate the square of the vector components
+    bool do_add_rms_partials;
+
+    // Cache most recent tensor that was converted into prealloc_y, and what pipeline it used to convert.
+    vk_pipeline_struct * prealloc_y_last_pipeline_used {};
+    const ggml_tensor * prealloc_y_last_tensor_used {};
+
+    // Track which nodes have been used since the last sync, and whether they were written to
+    std::vector<const ggml_tensor *> unsynced_nodes_written;
+    std::vector<const ggml_tensor *> unsynced_nodes_read;
+    // Track which prealloc buffers have pending reads that need to be synchronized.
+    // These are checked before writing to the buffer (and call ggml_vk_sync_buffers if set),
+    // and set to true after the buffer contents are consumed.
+    bool prealloc_x_need_sync, prealloc_y_need_sync, prealloc_split_k_need_sync;
 
     vk_buffer buffer_pool[MAX_VK_BUFFERS];
 
@@ -1401,7 +1503,7 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
 
     {
         std::lock_guard<std::recursive_mutex> guard(device->mutex);
-        device->pipelines.insert({ pipeline->name, pipeline });
+        device->all_pipelines.push_back(pipeline);
     }
 
     {
@@ -1838,14 +1940,18 @@ static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) {
     return { buf, 0, VK_WHOLE_SIZE };
 }
 
-static void ggml_vk_sync_buffers(vk_context& ctx) {
+static void ggml_vk_sync_buffers(ggml_backend_vk_context* ctx, vk_context& subctx) {
     VK_LOG_DEBUG("ggml_vk_sync_buffers()");
 
-    const bool transfer_queue = ctx->p->q->transfer_only;
+    const bool transfer_queue = subctx->p->q->transfer_only;
 
-    ctx->s->buffer.pipelineBarrier(
-        ctx->p->q->stage_flags,
-        ctx->p->q->stage_flags,
+    if (ctx) {
+        ctx->prealloc_x_need_sync = ctx->prealloc_y_need_sync = ctx->prealloc_split_k_need_sync = false;
+    }
+
+    subctx->s->buffer.pipelineBarrier(
+        subctx->p->q->stage_flags,
+        subctx->p->q->stage_flags,
         {},
         { {
           { !transfer_queue ? (vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) : (vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) },
@@ -1872,47 +1978,12 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
     );
 }
 
-enum FaCodePath {
-    FA_SCALAR,
-    FA_COOPMAT1,
-    FA_COOPMAT2,
-};
-
-static FaHeadSizes fa_get_head_sizes(uint32_t hsk, uint32_t hsv) {
-    if (hsk != 192 && hsk != 576 && hsk != hsv) {
-        return FA_HEAD_SIZE_UNSUPPORTED;
-    }
-    switch (hsk) {
-    case 64: return FA_HEAD_SIZE_64;
-    case 80: return FA_HEAD_SIZE_80;
-    case 96: return FA_HEAD_SIZE_96;
-    case 112: return FA_HEAD_SIZE_112;
-    case 128: return FA_HEAD_SIZE_128;
-    case 192:
-        if (hsv == 192) {
-            return FA_HEAD_SIZE_192;
-        } else if (hsv == 128) {
-            return FA_HEAD_SIZE_192_128;
-        } else {
-            return FA_HEAD_SIZE_UNSUPPORTED;
-        }
-    case 256: return FA_HEAD_SIZE_256;
-    case 576:
-        if (hsv == 512) {
-            return FA_HEAD_SIZE_576_512;
-        } else {
-            return FA_HEAD_SIZE_UNSUPPORTED;
-        }
-    default: return FA_HEAD_SIZE_UNSUPPORTED;
-    }
-}
-
 // number of rows/cols for flash attention shader
 static constexpr uint32_t flash_attention_num_small_rows = 32;
 static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
 
 static uint32_t get_fa_scalar_num_large_rows(uint32_t hsv) {
-    if (hsv >= 512) {
+    if (hsv >= 192) {
         return 2;
     } else {
         return 8;
@@ -1942,7 +2013,13 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
         if (small_rows) {
             return {scalar_flash_attention_num_small_rows, 64};
         } else {
-            return {get_fa_scalar_num_large_rows(hsv), 32};
+            if ((hsv | hsk) & 8) {
+                // HSV/HSK not being a multiple of 16 makes D_split smaller, which makes cols_per_iter
+                // larger, and Bc needs to be >= cols_per_thread. 64 is large enough, 32 is not.
+                return {get_fa_scalar_num_large_rows(hsv), 64};
+            } else {
+                return {get_fa_scalar_num_large_rows(hsv), 32};
+            }
         }
     }
 
@@ -1960,8 +2037,8 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
     }
 
     // small cols to reduce register count
-    if (ggml_is_quantized(type) || hsk >= 256) {
-        if (hsk >= 512) {
+    if (ggml_is_quantized(type) || hsk >= 256 || hsv >= 256) {
+        if (hsk >= 512 || hsv >= 512) {
             return {32, 32};
         } else {
             return {64, 32};
@@ -1970,6 +2047,10 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint3
     return {64, 64};
 }
 
+static uint32_t fa_align(FaCodePath path, uint32_t hsk, uint32_t hsv, ggml_type type, bool small_rows) {
+    return fa_rows_cols(path, hsk, hsv, 0, type, small_rows)[1];
+}
+
 static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector<uint32_t>& warptile, bool mul_mat_id, ggml_type src0_type) {
 
     uint32_t lut_size = 0;
@@ -2133,9 +2214,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
         s_mmq_wg_denoms_k = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul_id
-        l_warptile_mmqid = { 256, 128, 128, 16, 0 };
-        m_warptile_mmqid = { 256, 128, 64, 16, 0 };
-        s_warptile_mmqid = { 256, 128, 64, 16, 0 };
+        l_warptile_mmqid = { 256, 128, 128, 16, 0, device->subgroup_size };
+        m_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
+        s_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
         l_mmqid_wg_denoms = { 128, 128, 1 };
         m_mmqid_wg_denoms = { 128, 64, 1 };
         s_mmqid_wg_denoms = { 128, 64, 1 };
@@ -2235,11 +2316,14 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         if (!pipeline) {
             pipeline = std::make_shared<vk_pipeline_struct>();
+        }
+        if (!pipeline->initialized) {
             pipeline->name = name;
             pipeline->parameter_count = parameter_count;
             pipeline->push_constant_size = push_constant_size;
             pipeline->wg_denoms = wg_denoms;
             pipeline->align = align;
+            pipeline->initialized = true;
         }
 
         if (!pipeline->needed || pipeline->compiled) {
@@ -2285,26 +2369,30 @@ static void ggml_vk_load_shaders(vk_device& device) {
         return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split};
     };
 
-#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, HSK, HSV, HEAD_SIZES) \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-
 #define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64, 64, 64) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80, 80, 80) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96, 96, 96) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112, 112, 112) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128, 128, 128) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 192, 192, 192) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 192, 128, 192_128) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256, 256, 256) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 576, 512, 576_512)
+        for (auto &fa : device->pipeline_flash_attn_f32_f16[TYPE]) { \
+            uint32_t HSK = fa.first.HSK; \
+            uint32_t HSV = fa.first.HSV; \
+            bool small_rows = fa.first.small_rows; \
+            FaCodePath path = fa.first.path; \
+            bool aligned = fa.first.aligned; \
+            bool f32acc = fa.first.f32acc; \
+            if (path == FAPATH) { \
+                if (aligned) { \
+                    if (f32acc) { \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_align(FAPATH,HSK,HSV,TYPE,small_rows), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                    } else { \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows), fa_align(FAPATH,HSK,HSV,TYPE,small_rows), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                    } \
+                } else { \
+                    if (f32acc) { \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                    } else { \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                    } \
+                } \
+            } \
+        }
 
     CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, )
     CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, )
@@ -2327,7 +2415,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, FA_COOPMAT2, _cm2)
     }
 #endif
-#undef CREATE_FA2
 #undef CREATE_FA
 
 #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
@@ -2380,26 +2467,26 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
         }
 #endif
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
 #undef CREATE_MM
 #undef CREATE_MM2
     } else
@@ -2495,51 +2582,27 @@ static void ggml_vk_load_shaders(vk_device& device) {
         }
 #endif
 
-        if (device->coopmat_acc_f16_support) {
-            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-
-            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        } else {
-            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-
-            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        }
+        CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 #undef CREATE_MM2
 #undef CREATE_MM
     } else
@@ -2624,27 +2687,27 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
 
-        CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-
-        CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 #undef CREATE_MM2
 #undef CREATE_MMQ
 #undef CREATE_MM
@@ -2773,54 +2836,61 @@ static void ggml_vk_load_shaders(vk_device& device) {
         rm_stdq = 2;
     uint32_t rm_iq = 2 * rm_kq;
 
-    for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f32_f32_len, mul_mat_vec_bf16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f32_f32_len,   mul_mat_vec_iq1_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f32_f32_len,   mul_mat_vec_iq1_m_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f32_f32_len,  mul_mat_vec_iq2_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f32_f32_len,   mul_mat_vec_iq2_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f32_f32_len,   mul_mat_vec_iq3_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f32_f32_len,  mul_mat_vec_iq4_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f32_f32_len,  mul_mat_vec_iq4_nl_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32_"+std::to_string(i+1),   mul_mat_vec_mxfp4_f32_f32_len,   mul_mat_vec_mxfp4_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f16_f32_len, mul_mat_vec_bf16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_s_f16_f32_len,   mul_mat_vec_iq1_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq1_m_f16_f32_len,   mul_mat_vec_iq1_m_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f16_f32_len,  mul_mat_vec_iq2_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f16_f32_len,   mul_mat_vec_iq2_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f16_f32_len,   mul_mat_vec_iq3_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_xs_f16_f32_len,  mul_mat_vec_iq4_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f16_f32_len,  mul_mat_vec_iq4_nl_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32_"+std::to_string(i+1),   mul_mat_vec_mxfp4_f16_f32_len,   mul_mat_vec_mxfp4_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true);
+    for (uint32_t w = 0; w < DMMV_WG_SIZE_COUNT; ++w) {
+        uint32_t wg_size_subgroup16 = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_16 : (subgroup_size_16 * 4);
+        uint32_t wg_size_subgroup   = (w == DMMV_WG_SIZE_SUBGROUP) ? device->subgroup_size : (device->subgroup_size * 4);
+
+        const bool s = device->subgroup_add && device->architecture != vk_device_architecture::AMD_GCN;
+
+        for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32",  arr_dmmv_f32_f32_f32_len[s],  arr_dmmv_f32_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32",  arr_dmmv_f16_f32_f32_len[s],  arr_dmmv_f16_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32", arr_dmmv_bf16_f32_f32_len[s], arr_dmmv_bf16_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32", arr_dmmv_q4_0_f32_f32_len[s], arr_dmmv_q4_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32", arr_dmmv_q4_1_f32_f32_len[s], arr_dmmv_q4_1_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32", arr_dmmv_q5_0_f32_f32_len[s], arr_dmmv_q5_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32", arr_dmmv_q5_1_f32_f32_len[s], arr_dmmv_q5_1_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[s], arr_dmmv_q8_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[s], arr_dmmv_q2_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[s], arr_dmmv_q3_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[s], arr_dmmv_q4_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[s], arr_dmmv_q5_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32", arr_dmmv_q6_k_f32_f32_len[s], arr_dmmv_q6_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32",   arr_dmmv_iq1_s_f32_f32_len[s],   arr_dmmv_iq1_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32",   arr_dmmv_iq1_m_f32_f32_len[s],   arr_dmmv_iq1_m_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32", arr_dmmv_iq2_xxs_f32_f32_len[s], arr_dmmv_iq2_xxs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32",  arr_dmmv_iq2_xs_f32_f32_len[s],  arr_dmmv_iq2_xs_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32",   arr_dmmv_iq2_s_f32_f32_len[s],   arr_dmmv_iq2_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32", arr_dmmv_iq3_xxs_f32_f32_len[s], arr_dmmv_iq3_xxs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32",   arr_dmmv_iq3_s_f32_f32_len[s],   arr_dmmv_iq3_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32",  arr_dmmv_iq4_xs_f32_f32_len[s],  arr_dmmv_iq4_xs_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32",  arr_dmmv_iq4_nl_f32_f32_len[s],  arr_dmmv_iq4_nl_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32",   arr_dmmv_mxfp4_f32_f32_len[s],   arr_dmmv_mxfp4_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32",  arr_dmmv_f32_f16_f32_len[s],  arr_dmmv_f32_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32",  arr_dmmv_f16_f16_f32_len[s],  arr_dmmv_f16_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32", arr_dmmv_bf16_f16_f32_len[s], arr_dmmv_bf16_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32", arr_dmmv_q4_0_f16_f32_len[s], arr_dmmv_q4_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32", arr_dmmv_q4_1_f16_f32_len[s], arr_dmmv_q4_1_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32", arr_dmmv_q5_0_f16_f32_len[s], arr_dmmv_q5_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32", arr_dmmv_q5_1_f16_f32_len[s], arr_dmmv_q5_1_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[s], arr_dmmv_q8_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[s], arr_dmmv_q2_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[s], arr_dmmv_q3_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[s], arr_dmmv_q4_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[s], arr_dmmv_q5_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32", arr_dmmv_q6_k_f16_f32_len[s], arr_dmmv_q6_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32",   arr_dmmv_iq1_s_f16_f32_len[s],   arr_dmmv_iq1_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32",   arr_dmmv_iq1_m_f16_f32_len[s],   arr_dmmv_iq1_m_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32", arr_dmmv_iq2_xxs_f16_f32_len[s], arr_dmmv_iq2_xxs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32",  arr_dmmv_iq2_xs_f16_f32_len[s],  arr_dmmv_iq2_xs_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32",   arr_dmmv_iq2_s_f16_f32_len[s],   arr_dmmv_iq2_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32", arr_dmmv_iq3_xxs_f16_f32_len[s], arr_dmmv_iq3_xxs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32",   arr_dmmv_iq3_s_f16_f32_len[s],   arr_dmmv_iq3_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32",  arr_dmmv_iq4_xs_f16_f32_len[s],  arr_dmmv_iq4_xs_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32",  arr_dmmv_iq4_nl_f16_f32_len[s],  arr_dmmv_iq4_nl_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32",   arr_dmmv_mxfp4_f16_f32_len[s],   arr_dmmv_mxfp4_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+        }
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
@@ -2924,8 +2994,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_f32, "rms_norm_mul_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 1}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_f32, "rms_norm_mul_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 1}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_partials_f32, "rms_norm_partials_f32", rms_norm_partials_f32_len, rms_norm_partials_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_partials_f32, "rms_norm_mul_partials_f32", rms_norm_partials_f32_len, rms_norm_partials_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 1}, 1, true);
+
     ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
 
@@ -2995,22 +3069,31 @@ static void ggml_vk_load_shaders(vk_device& device) {
     };
 
     bool rte = device->float_controls_rte_fp16;
-#define CREATE_BINARY(name, namemod, spec) \
+#define CREATE_BINARY(name, namemod, spec, bindings) \
     for (int s0 : {0,1}) for (int s1 : {0,1}) for (int d : {0,1}) \
         ggml_vk_create_pipeline(device, device->pipeline_ ## name ## namemod[s0][s1][d], \
                                 #name + get_suffix(s0, s1, d) + #namemod, name ## _len[s0][s1][d][rte], name ## _data[s0][s1][d][rte], \
-                                "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1);
-
-    CREATE_BINARY(add, , {0})
-    CREATE_BINARY(add, _norepeat, {1})
-    CREATE_BINARY(sub, , {0})
-    CREATE_BINARY(sub, _norepeat, {1})
-    CREATE_BINARY(mul, , {0})
-    CREATE_BINARY(mul, _norepeat, {1})
-    CREATE_BINARY(div, , {0})
-    CREATE_BINARY(div, _norepeat, {1})
+                                "main", (bindings), sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1);
+
+    CREATE_BINARY(add, , {0}, 4)
+    CREATE_BINARY(add, _norepeat, {1}, 4)
+    CREATE_BINARY(sub, , {0}, 3)
+    CREATE_BINARY(sub, _norepeat, {1}, 3)
+    CREATE_BINARY(mul, , {0}, 3)
+    CREATE_BINARY(mul, _norepeat, {1}, 3)
+    CREATE_BINARY(div, , {0}, 3)
+    CREATE_BINARY(div, _norepeat, {1}, 3)
+    CREATE_BINARY(add_rms, , {0}, 4)
+    CREATE_BINARY(add_rms, _norepeat, {1}, 4)
 #undef CREATE_BINARY
 
+    if (device->multi_add) {
+        for (uint32_t i = 0; i < MAX_FUSED_ADDS; ++i) {
+            ggml_vk_create_pipeline(device, device->pipeline_multi_add[i],     "multi_add_f32_"     + std::to_string(i+1), multi_add_f32_len,     multi_add_f32_data,     "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1);
+            ggml_vk_create_pipeline(device, device->pipeline_multi_add_rms[i], "multi_add_rms_f32_" + std::to_string(i+1), multi_add_rms_f32_len, multi_add_rms_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1);
+        }
+    }
+
     ggml_vk_create_pipeline(device, device->pipeline_add_id_f32, "add_id_f32", add_id_f32_len, add_id_f32_data, "main", 4, sizeof(vk_op_add_id_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -3026,6 +3109,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_sqrt_f32, "sqrt_f32", sqrt_f32_len, sqrt_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -3042,6 +3126,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);  \
     ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
+    CREATE_UNARY(exp)
     CREATE_UNARY(gelu)
     CREATE_UNARY(gelu_erf)
     CREATE_UNARY(gelu_quick)
@@ -3096,11 +3181,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     }
 
-    ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
+    for (uint32_t i = 0; i < num_argsort_pipelines; ++i) {
+        ggml_vk_create_pipeline(device, device->pipeline_argsort_f32[i], "argsort_f32_"+std::to_string(i), argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1u<<i, 1, 1}, {1u<<i, i}, 1, true);
+    }
 
     ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1);
 
@@ -3123,6 +3210,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_opt_step_sgd_f32, "opt_step_sgd_f32", opt_step_sgd_f32_len, opt_step_sgd_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+
     // conv2d
     for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
         uint32_t conv2d_WG_SIZE  = 256;
@@ -3227,6 +3316,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f32, "conv2d_dw_cwhn_f32", conv2d_dw_cwhn_f32_len, conv2d_dw_cwhn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f16_f32, "conv2d_dw_whcn_f16_f32", conv2d_dw_whcn_f16_f32_len, conv2d_dw_whcn_f16_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f16_f32, "conv2d_dw_cwhn_f16_f32", conv2d_dw_cwhn_f16_f32_len, conv2d_dw_cwhn_f16_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
 
     for (auto &c : compiles) {
         c.wait();
@@ -3548,6 +3639,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
 
+        device->multi_add = vk12_props.shaderRoundingModeRTEFloat16 &&
+                            device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_multi_add_push_constants) &&
+                            vk12_features.runtimeDescriptorArray &&
+                            device->vendor_id != VK_VENDOR_ID_INTEL &&
+                            getenv("GGML_VK_DISABLE_MULTI_ADD") == nullptr;
+
         if (device->subgroup_size_control) {
             device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
             device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
@@ -3861,6 +3958,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->disable_fusion = getenv("GGML_VK_DISABLE_FUSION") != nullptr;
 
+        device->add_rms_fusion = !device->disable_fusion &&
+                                 device->subgroup_add &&
+                                 device->vendor_id != VK_VENDOR_ID_INTEL;
+        device->partials_binding_alignment =
+            std::max(4u, (uint32_t)device->properties.limits.minStorageBufferOffsetAlignment);
+
         return device;
     }
 
@@ -4372,7 +4475,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     return (ctx->device->fp16 && prec == GGML_PREC_DEFAULT) ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
 }
 
-static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols) {
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols, uint32_t m, uint32_t k) {
     VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
     GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
     GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols);
@@ -4406,7 +4509,24 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
             return nullptr;
     }
 
-    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type][num_cols-1];
+    // heuristic to choose workgroup size
+    uint32_t dmmv_wg = DMMV_WG_SIZE_SUBGROUP;
+    if (ctx->device->vendor_id == VK_VENDOR_ID_NVIDIA || ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
+        // Prefer larger workgroups when M is small, to spread the work out more
+        // and keep more SMs busy.
+        // q6_k seems to prefer small workgroup size even for "medium" values of M.
+        if (a_type == GGML_TYPE_Q6_K) {
+            if (m < 4096 && k >= 1024) {
+                dmmv_wg = DMMV_WG_SIZE_LARGE;
+            }
+        } else {
+            if (m <= 8192 && k >= 1024) {
+                dmmv_wg = DMMV_WG_SIZE_LARGE;
+            }
+        }
+    }
+
+    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[dmmv_wg][a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[dmmv_wg][a_type][num_cols-1];
 }
 
 static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
@@ -4461,7 +4581,17 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
             return nullptr;
     }
 
-    return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
+    // XXX TODO 'prec' is not actually allowed in mul_mat_id.
+    bool prefer_fp16acc = ctx->device->fp16 /*&& prec == GGML_PREC_DEFAULT*/;
+    bool support_fp16acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc != nullptr;
+    bool support_fp32acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc != nullptr;
+
+    if (support_fp16acc && (prefer_fp16acc || !support_fp32acc)) {
+        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc;
+    } else {
+        GGML_ASSERT(support_fp32acc);
+        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
+    }
 }
 
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
@@ -4800,7 +4930,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
             }
         }
 
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(ctx, subctx);
         subctx->s->buffer.copyBuffer(buf->buffer, dst->buffer, slices);
         return;
     }
@@ -4815,7 +4945,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
     ggml_vk_ensure_sync_staging_buffer(ctx->device, copy_size);
     VkBufferCopy buf_copy{ 0, offset, copy_size };
 
-    ggml_vk_sync_buffers(subctx);
+    ggml_vk_sync_buffers(ctx, subctx);
     vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
 
     for (uint64_t i3 = 0; i3 < ne3; i3++) {
@@ -4869,7 +4999,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
             }
         }
 
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(nullptr, subctx);
         subctx->s->buffer.copyBuffer(buf->buffer, dst->buffer, slices);
         return;
     }
@@ -4890,7 +5020,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
         offset,
         copy_size};
 
-    ggml_vk_sync_buffers(subctx);
+    ggml_vk_sync_buffers(nullptr, subctx);
     vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging_buffer->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
 
     if (width == spitch) {
@@ -4970,7 +5100,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
 
     if (buf != nullptr) {
         // Memory is pinned, use as staging buffer
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(nullptr, subctx);
         subctx->s->buffer.copyBuffer(src->buffer, buf->buffer, slices);
 
         return;
@@ -4987,7 +5117,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
 
     vk_buffer& staging_buffer = src->device->sync_staging;
 
-    ggml_vk_sync_buffers(subctx);
+    ggml_vk_sync_buffers(nullptr, subctx);
     subctx->s->buffer.copyBuffer(src->buffer, staging_buffer->buffer, slices);
 
     deferred_memcpy(dst, staging_buffer->ptr, copy_size, &subctx->out_memcpys);
@@ -5177,13 +5307,16 @@ static void ggml_vk_matmul(
         uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3,
         uint32_t padded_n) {
         VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", padded_n: " << padded_n << ")");
-    ggml_vk_sync_buffers(subctx);
     if (split_k == 1) {
         const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n };
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc, { m, n, batch });
         return;
     }
 
+    if (ctx->prealloc_split_k_need_sync) {
+        ggml_vk_sync_buffers(ctx, subctx);
+    }
+
     GGML_ASSERT(batch_stride_d == m * n);
 
     // Round the split size up to a multiple of 256 (k-quant alignment)
@@ -5193,9 +5326,10 @@ static void ggml_vk_matmul(
     const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k_split, ne02, ne12, broadcast2, broadcast3, padded_n };
     // Make sure enough workgroups get assigned for split k to work
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
-    ggml_vk_sync_buffers(subctx);
+    ggml_vk_sync_buffers(ctx, subctx);
     const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2, { m * n * batch, 1, 1 });
+    ctx->prealloc_split_k_need_sync = true;
 }
 
 static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, uint32_t m, uint32_t n, bool aligned, ggml_type src0_type) {
@@ -5240,7 +5374,6 @@ static void ggml_vk_matmul_id(
         "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " <<
         "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " <<
         "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")");
-    ggml_vk_sync_buffers(subctx);
     const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d,
                                               nei0, nei1, nbi1, ne11, padded_n };
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, pc, { m, nei1, n_as });
@@ -5371,8 +5504,8 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     };
     init_pushconst_fastdiv(pc);
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, pc, elements);
+    ggml_vk_sync_buffers(ctx, subctx);
 }
 
 static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
@@ -5390,8 +5523,8 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
 
     vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
 
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
+    ggml_vk_sync_buffers(ctx, subctx);
 }
 
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5586,18 +5719,39 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         GGML_ASSERT(qy_sz == y_sz);
     }
 
+    if (x_non_contig || qx_needs_dequant) {
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    if (y_non_contig || quantize_y) {
+        if (ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+
     if (x_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_sync_buffers(ctx, subctx);
     }
     if (y_non_contig) {
-        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+        if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
+            ctx->prealloc_y_last_tensor_used != src1) {
+            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+            ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
+            ctx->prealloc_y_last_tensor_used = src1;
+        }
     }
     if (quantize_y) {
-        ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13);
+        if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
+            ctx->prealloc_y_last_tensor_used != src1) {
+            ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13);
+            ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
+            ctx->prealloc_y_last_tensor_used = src1;
+        }
     }
 
     uint32_t stride_batch_x = ne00*ne01;
@@ -5620,6 +5774,13 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21,
         split_k, ne12*ne13, ne02, ne12, r2, r3, padded_n
     );  // NOLINT
+
+    if (x_non_contig || qx_needs_dequant) {
+        ctx->prealloc_x_need_sync = true;
+    }
+    if (y_non_contig || quantize_y) {
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5703,7 +5864,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
-    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11);
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11, ne20, ne00);
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
@@ -5766,13 +5927,29 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         GGML_ASSERT(qy_sz == y_sz);
     }
 
+    if (x_non_contig) {
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    if (y_non_contig) {
+        if (ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+
     if (x_non_contig) {
         GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
     }
     if (y_non_contig) {
         GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
-        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+        if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
+            ctx->prealloc_y_last_tensor_used != src1) {
+            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+            ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
+            ctx->prealloc_y_last_tensor_used = src1;
+        }
     }
 
     // For batch_n, the A matrix is the same for each batch, and B/D use the row stride as the batch stride
@@ -5804,10 +5981,16 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         stride_batch_x, stride_batch_y, stride_batch_d,
         (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
     };
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
                               { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
                               pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
+
+    if (x_non_contig) {
+        ctx->prealloc_x_need_sync = true;
+    }
+    if (y_non_contig) {
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5894,7 +6077,6 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
         workgroups_z /= gqa_ratio;
     }
 
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, workgroups_z });
 }
 
@@ -5981,7 +6163,6 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     // compute
     const std::array<uint32_t, 12> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)), nb03, nb13, nb23 };
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
         { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { (uint32_t)ne03, (uint32_t)ne01, (uint32_t)ne12 });
 }
@@ -6193,16 +6374,32 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         GGML_ASSERT(qy_sz == y_sz);
     }
 
+    if (x_non_contig || qx_needs_dequant) {
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    if (y_non_contig) {
+        if (ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+
     if (x_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
             { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_sync_buffers(ctx, subctx);
     }
     if (y_non_contig) {
-        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+        if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
+            ctx->prealloc_y_last_tensor_used != src1) {
+            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+            ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
+            ctx->prealloc_y_last_tensor_used = src1;
+        }
     }
 
     uint32_t stride_batch_x = ne00*ne01;
@@ -6225,6 +6422,13 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         stride_batch_x, stride_batch_y, ne20*ne21,
         n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n
     );  // NOLINT
+
+    if (x_non_contig || qx_needs_dequant) {
+        ctx->prealloc_x_need_sync = true;
+    }
+    if (y_non_contig) {
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, bool dryrun = false) {
@@ -6384,13 +6588,29 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         GGML_ASSERT(qy_sz == y_sz);
     }
 
+    if (x_non_contig) {
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+    if (y_non_contig) {
+        if (ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+    }
+
     if (x_non_contig) {
         GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
     }
     if (y_non_contig) {
         GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
-        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+        if (ctx->prealloc_y_last_pipeline_used != to_fp16_vk_1.get() ||
+            ctx->prealloc_y_last_tensor_used != src1) {
+            ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+            ctx->prealloc_y_last_pipeline_used = to_fp16_vk_1.get();
+            ctx->prealloc_y_last_tensor_used = src1;
+        }
     }
 
     uint32_t stride_batch_y = ne10*ne11;
@@ -6415,11 +6635,17 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         (uint32_t)x_ne, stride_batch_y, (uint32_t)(ne20*ne21),
         (uint32_t)nei0, (uint32_t)ne11,
     };
-    ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
         { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
         vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } },
         pc, { groups_x, (uint32_t)nei0, groups_z });
+
+    if (x_non_contig) {
+        ctx->prealloc_x_need_sync = true;
+    }
+    if (y_non_contig) {
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
@@ -6434,22 +6660,29 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
         GGML_ASSERT(nei0 <= 4096);
         const uint32_t split_size = std::min(nei1, 4096u / nei0);
 
-        ggml_tensor src1_copy = *src1;
-        ggml_tensor src2_copy = *src2;
-        ggml_tensor dst_copy = *dst;
+        if (split_size == nei1) {
+            ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
+        } else {
+            ggml_tensor src1_copy = *src1;
+            ggml_tensor src2_copy = *src2;
+            ggml_tensor dst_copy = *dst;
 
-        for (uint32_t token_start = 0; token_start < nei1; token_start += split_size) {
-            const uint32_t n_tokens = std::min(split_size, nei1 - token_start);
+            for (uint32_t token_start = 0; token_start < nei1; token_start += split_size) {
+                const uint32_t n_tokens = std::min(split_size, nei1 - token_start);
 
-            src1_copy.view_offs = src1->view_offs + token_start * src1_copy.nb[2];
-            src2_copy.view_offs = src2->view_offs + token_start * src2_copy.nb[1];
-            dst_copy.view_offs = dst->view_offs + token_start * dst_copy.nb[2];
+                src1_copy.view_offs = src1->view_offs + token_start * src1_copy.nb[2];
+                src2_copy.view_offs = src2->view_offs + token_start * src2_copy.nb[1];
+                dst_copy.view_offs = dst->view_offs + token_start * dst_copy.nb[2];
 
-            src1_copy.ne[2] = n_tokens;
-            src2_copy.ne[1] = n_tokens;
-            dst_copy.ne[2] = n_tokens;
+                src1_copy.ne[2] = n_tokens;
+                src2_copy.ne[1] = n_tokens;
+                dst_copy.ne[2] = n_tokens;
 
-            ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, &src1_copy, &src2_copy, &dst_copy, dryrun);
+                ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, &src1_copy, &src2_copy, &dst_copy, dryrun);
+                // invalidate cached prealloc_y, can't cache based on the copy of the ggml_tensor
+                ctx->prealloc_y_last_pipeline_used = {};
+                ctx->prealloc_y_last_tensor_used = nullptr;
+            }
         }
     }
 }
@@ -6483,18 +6716,21 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     const uint32_t Br = coopmat1_flash_attention_num_large_rows;
     const uint32_t Bc = scalar_flash_attention_Bc;
 
+    const uint32_t hsk_pad = ROUNDUP_POW2(hsk, 16);
+
     const uint32_t acctype = f32acc ? 4 : 2;
     const uint32_t f16vec4 = 8;
 
     const uint32_t tmpsh = wg_size * sizeof(float);
     const uint32_t tmpshv4 = wg_size * 4 * acctype;
 
-    const uint32_t Qf = Br * (hsk / 4 + 2) * f16vec4;
+    const uint32_t qstride = hsk_pad / 4 + 2;
+    const uint32_t Qf = Br * qstride * f16vec4;
 
     const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br;
     const uint32_t sfsh = Bc * sfshstride * acctype;
 
-    const uint32_t kshstride = hsk / 4 + 2;
+    const uint32_t kshstride = hsk_pad / 4 + 2;
     const uint32_t ksh = Bc * kshstride * f16vec4;
 
     const uint32_t slope = Br * sizeof(float);
@@ -6605,7 +6841,6 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         workgroups_y /= N;
     }
 
-    vk_pipeline *pipelines;
     bool small_rows = N <= get_fa_num_small_rows(path);
 
     // coopmat1 does not actually support "small rows" (it needs 16 rows).
@@ -6625,37 +6860,36 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         small_rows = true;
     }
 
-    bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
-
-    FaHeadSizes head_sizes = fa_get_head_sizes(k->ne[0], v->ne[0]);
-
-    switch (path) {
-    case FA_SCALAR:
-        pipelines = &ctx->device->pipeline_flash_attn_f32_f16[k->type][head_sizes][f32acc][small_rows][0];
-        break;
-    case FA_COOPMAT1:
-        pipelines = &ctx->device->pipeline_flash_attn_f32_f16_cm1[k->type][head_sizes][f32acc][small_rows][0];
-        break;
-    case FA_COOPMAT2:
-        pipelines = &ctx->device->pipeline_flash_attn_f32_f16_cm2[k->type][head_sizes][f32acc][small_rows][0];
-        break;
-    default:
-        GGML_ASSERT(0);
-    }
-    assert(pipelines);
-
     const uint32_t q_stride = (uint32_t)(nbq1 / ggml_type_size(q->type));
     const uint32_t k_stride = (uint32_t)(nbk1 / ggml_type_size(k->type));
     const uint32_t v_stride = (uint32_t)(nbv1 / ggml_type_size(v->type));
 
-    bool aligned = (KV % pipelines[1]->align) == 0 &&
+    uint32_t alignment = fa_align(path, HSK, HSV, k->type, small_rows);
+    bool aligned = (KV % alignment) == 0 &&
                    // the "aligned" shader variant will forcibly align strides, for performance
                    (q_stride & 7) == 0 && (k_stride & 7) == 0 && (v_stride & 7) == 0;
 
+    // Need to use the coopmat2 variant that clamps loads when HSK/HSV aren't sufficiently aligned.
+    if (((HSK | HSV) % 16) != 0 && path == FA_COOPMAT2) {
+        aligned = false;
+    }
     // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads
     GGML_ASSERT((nem1 % GGML_KQ_MASK_PAD) == 0);
 
-    vk_pipeline pipeline = pipelines[aligned];
+    bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
+
+    vk_fa_pipeline_state fa_pipeline_state(HSK, HSV, small_rows, path, aligned, f32acc);
+
+    vk_pipeline pipeline = nullptr;
+
+    auto &pipelines = ctx->device->pipeline_flash_attn_f32_f16[k->type];
+    auto it = pipelines.find(fa_pipeline_state);
+    if (it != pipelines.end()) {
+        pipeline = it->second;
+    } else {
+        pipelines[fa_pipeline_state] = pipeline = std::make_shared<vk_pipeline_struct>();
+    }
+
     assert(pipeline);
 
     uint32_t split_kv = KV;
@@ -6671,7 +6905,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         if (split_k > 1) {
             // Try to evenly split KV into split_k chunks, but it needs to be a multiple
             // of "align", so recompute split_k based on that.
-            split_kv = ROUNDUP_POW2(std::max(1u, KV / split_k), pipelines[1]->align);
+            split_kv = ROUNDUP_POW2(std::max(1u, KV / split_k), alignment);
             split_k = CEIL_DIV(KV, split_kv);
             workgroups_x = split_k;
         }
@@ -6795,9 +7029,11 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                               mask_n_head_log2, m0, m1,
                                               gqa_ratio, split_kv, split_k };
 
-    ggml_vk_sync_buffers(subctx);
-
     if (split_k > 1) {
+        if (ctx->prealloc_split_k_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {
                                         vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE},
@@ -6813,7 +7049,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     // cancel out the divide by wg_denoms[0].
                                     pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
 
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(ctx, subctx);
         const std::array<uint32_t, 5> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k, (sinks != nullptr) };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
                                     {
@@ -6822,6 +7058,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                         vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                     },
                                     pc2, { (uint32_t)ne1, HSV, (uint32_t)ne3 });
+        ctx->prealloc_split_k_need_sync = true;
     } else {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {
@@ -6864,7 +7101,7 @@ static std::array<uint32_t, 3> ggml_vk_get_conv_elements(const ggml_tensor *dst)
     return elements;
 }
 
-static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) {
+static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * dst, ggml_op op) {
     switch (op) {
     case GGML_OP_GET_ROWS:
         GGML_ASSERT(src1->type == GGML_TYPE_I32);
@@ -6892,8 +7129,20 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         switch (op) {
         case GGML_OP_ADD:
         {
-            auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_norepeat : ctx->device->pipeline_add;
-            return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16];
+            if (ctx->num_additional_fused_ops > 0) {
+                if (ctx->do_add_rms_partials) {
+                    return ctx->device->pipeline_multi_add_rms[ctx->num_additional_fused_ops];
+                } else {
+                    return ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+                }
+            }
+            if (ctx->do_add_rms_partials) {
+                auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_rms_norepeat : ctx->device->pipeline_add_rms;
+                return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16];
+            } else {
+                auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_norepeat : ctx->device->pipeline_add;
+                return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16];
+            }
         }
         case GGML_OP_SUB:
         {
@@ -6953,6 +7202,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_sqr_f32;
         }
         return nullptr;
+    case GGML_OP_SQRT:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_sqrt_f32;
+        }
+        return nullptr;
     case GGML_OP_SIN:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_sin_f32;
@@ -7011,7 +7265,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         return nullptr;
     case GGML_OP_RMS_NORM:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->num_additional_fused_ops > 0 ? ctx->device->pipeline_rms_norm_mul_f32 : ctx->device->pipeline_rms_norm_f32;
+            if (ctx->do_add_rms_partials) {
+                return ctx->num_additional_fused_ops > 0 ? ctx->device->pipeline_rms_norm_mul_partials_f32 : ctx->device->pipeline_rms_norm_partials_f32;
+            } else {
+                return ctx->num_additional_fused_ops > 0 ? ctx->device->pipeline_rms_norm_mul_f32 : ctx->device->pipeline_rms_norm_f32;
+            }
         }
         return nullptr;
     case GGML_OP_RMS_NORM_BACK:
@@ -7032,6 +7290,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
 
         switch (ggml_get_unary_op(dst)) {
+            case GGML_UNARY_OP_EXP:
+                return ctx->device->pipeline_exp[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_SILU:
                 return ctx->device->pipeline_silu[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_GELU:
@@ -7136,11 +7396,13 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         }
     case GGML_OP_ARGSORT:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) {
-            return ctx->device->pipeline_argsort_f32;
+            uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
+            return ctx->device->pipeline_argsort_f32[idx];
         }
         return nullptr;
     case GGML_OP_SUM:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_sum_rows_f32;
         }
@@ -7193,6 +7455,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_opt_step_adamw_f32;
         }
         return nullptr;
+    case GGML_OP_OPT_STEP_SGD:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_opt_step_sgd_f32;
+        }
+        return nullptr;
     case GGML_OP_LEAKY_RELU:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_leaky_relu_f32;
@@ -7235,6 +7502,12 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             } else if (ggml_is_contiguous_channels(src1)) {
                 return ctx->device->pipeline_conv2d_dw_cwhn_f32;
             }
+        } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+            if (ggml_is_contiguous(src1)) {
+                return ctx->device->pipeline_conv2d_dw_whcn_f16_f32;
+            } else if (ggml_is_contiguous_channels(src1)) {
+                return ctx->device->pipeline_conv2d_dw_cwhn_f16_f32;
+            }
         }
         return nullptr;
     default:
@@ -7256,6 +7529,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
@@ -7267,6 +7541,9 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_IM2COL:
     case GGML_OP_SET_ROWS:
+    case GGML_OP_SUM:
+    case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
         return true;
     default:
         return false;
@@ -7301,6 +7578,16 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
     GGML_UNUSED(src2);
 }
 
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_sum_rows_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    p.misalign_offsets = (a_offset << 16) | d_offset;
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+}
+
 template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
     const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
     const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
@@ -7451,10 +7738,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
 
     if (op_supports_incontiguous) {
-        x_sz = ggml_nbytes(src0);
-        y_sz = use_src1 ? ggml_nbytes(src1) : 0;
-        z_sz = use_src2 ? ggml_nbytes(src2) : 0;
-        d_sz = ggml_nbytes(dst);
+        x_sz = ggml_nbytes(src0) + get_misalign_bytes(ctx, src0);
+        y_sz = use_src1 ? ggml_nbytes(src1) + get_misalign_bytes(ctx, src1) : 0;
+        z_sz = use_src2 ? ggml_nbytes(src2) + get_misalign_bytes(ctx, src2) : 0;
+        d_sz = ggml_nbytes(dst) + get_misalign_bytes(ctx, dst);
 
         if (x_buf_offset + x_sz >= d_X->size) {
             x_sz = VK_WHOLE_SIZE;
@@ -7482,6 +7769,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_SOFT_MAX:
     case GGML_OP_SOFT_MAX_BACK:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
     case GGML_OP_ARGMAX:
         {
             const uint32_t nr = ggml_nrows(src0);
@@ -7494,7 +7782,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             }
         } break;
     case GGML_OP_RMS_NORM:
-        elements = { (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne03 };
+        if (ctx->do_add_rms_partials) {
+            // Run one element per thread, 128 threads per workgroup
+            elements = { (uint32_t)CEIL_DIV(ne00, 128), 1, 1 };
+        } else {
+            elements = { (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne03 };
+        }
         break;
 
     case GGML_OP_SUM:
@@ -7561,6 +7854,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_MUL:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
@@ -7642,7 +7936,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         }
     }
 
-    if (op == GGML_OP_GLU) {
+    if (op == GGML_OP_ADD || op == GGML_OP_RMS_NORM) {
+        vk_buffer d_A = ctx->do_add_rms_partials ? ctx->prealloc_add_rms_partials : d_X;
+        size_t a_buf_offset = ctx->do_add_rms_partials ? ctx->prealloc_size_add_rms_partials_offset : 0;
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+            { vk_subbuffer{ d_X, x_buf_offset, x_sz },
+              vk_subbuffer{ d_Y, y_buf_offset, y_sz },
+              vk_subbuffer{ d_D, d_buf_offset, d_sz },
+              vk_subbuffer{ d_A, a_buf_offset, VK_WHOLE_SIZE },
+            }, pc, elements);
+    } else if (op == GGML_OP_GLU) {
         // Empty src1 is possible in glu, but the shader needs a buffer
         vk_subbuffer subbuf_y;
         if (use_src1) {
@@ -7651,7 +7954,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             subbuf_y = { d_X, 0, x_sz };
         }
 
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_SOFT_MAX) {
         // Empty src1 and src2 is possible in soft_max, but the shader needs a buffer
@@ -7669,7 +7971,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             subbuf_z = { d_X, 0, x_sz };
         }
 
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
         // Empty src2 is possible in rope, but the shader needs a buffer
@@ -7680,26 +7981,23 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             subbuf_z = { d_X, 0, x_sz };
         }
 
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_IM2COL) {
         // im2col uses only src1 and dst buffers
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_COUNT_EQUAL) {
-        ggml_vk_sync_buffers(subctx);
         // count_equal assumes that destination buffer is initialized with zeroes
         ggml_vk_buffer_memset_async(subctx, d_D, d_buf_offset, 0, d_sz);
-        ggml_vk_sync_buffers(subctx);
+        ggml_vk_sync_buffers(ctx, subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
+    } else if (op == GGML_OP_OPT_STEP_SGD) {
+        // OPT_STEP_SGD works on src0, it does not need dst
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz } }, pc, elements);
     } else if (use_src2) {
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (use_src1) {
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else {
-        ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     }
 }
@@ -7739,6 +8037,116 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
+static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
+    const ggml_tensor *first_node = cgraph->nodes[node_idx];
+    const ggml_tensor *dst = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
+
+    // Make a list of all the tensors used by the op.
+    // Last element of the list is the dest tensor.
+    const ggml_tensor *tensors[MAX_PARAMETER_COUNT];
+    uint32_t num_srcs = ctx->num_additional_fused_ops + 2;
+    uint32_t num_tensors = num_srcs + 1;
+    GGML_ASSERT(num_tensors + ctx->do_add_rms_partials <= MAX_PARAMETER_COUNT);
+
+    tensors[0] = first_node->src[0];
+    tensors[1] = first_node->src[1];
+    for (int32_t i = 0; i < ctx->num_additional_fused_ops; ++i) {
+        // check whether the previous result is src[0] or src[1]
+        if (cgraph->nodes[node_idx + i] == cgraph->nodes[node_idx + i + 1]->src[0]) {
+            tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[1];
+        } else {
+            tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[0];
+        }
+    }
+    tensors[num_srcs] = dst;
+
+    vk_op_multi_add_push_constants pc;
+    pc.ne20 = (uint32_t)dst->ne[0];
+    pc.ne21 = (uint32_t)dst->ne[1];
+    pc.ne22 = (uint32_t)dst->ne[2];
+    pc.ne23 = (uint32_t)dst->ne[3];
+
+    for (uint32_t i = 0; i < num_tensors; ++i) {
+        const ggml_tensor *t = tensors[i];
+        pc.nb[i][0] = (uint32_t)t->nb[0] / sizeof(float);
+        pc.nb[i][1] = (uint32_t)t->nb[1] / sizeof(float);
+        pc.nb[i][2] = (uint32_t)t->nb[2] / sizeof(float);
+        pc.nb[i][3] = (uint32_t)t->nb[3] / sizeof(float);
+    }
+    pc.rms_partials = ctx->do_add_rms_partials;
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, tensors[0], tensors[1], nullptr, dst, dst->op);
+
+    if (pipeline == nullptr) {
+        std::cerr << "ggml_vulkan: Error: Missing multi_add";
+        GGML_ABORT("fatal error");
+    }
+
+    if (dryrun) {
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+        return;
+    }
+
+    ggml_backend_vk_buffer_context * buf_ctx[MAX_PARAMETER_COUNT];
+    vk_buffer buf[MAX_PARAMETER_COUNT];
+    size_t offset[MAX_PARAMETER_COUNT];
+    bool uma[MAX_PARAMETER_COUNT];
+
+    for (uint32_t i = 0; i < num_tensors; ++i) {
+        buf_ctx[i] = (ggml_backend_vk_buffer_context *)tensors[i]->buffer->context;
+        buf[i] = nullptr;
+        offset[i] = 0;
+        uma[i] = false;
+
+        if (ctx->device->uma) {
+            ggml_vk_host_get(ctx->device, tensors[i]->data, buf[i], offset[i]);
+            uma[i] = buf[i] != nullptr;
+        }
+        if (!uma[i]) {
+            buf[i] = buf_ctx[i]->dev_buffer;
+            offset[i] = vk_tensor_offset(tensors[i]) + tensors[i]->view_offs;
+        }
+        GGML_ASSERT(buf[i] != nullptr);
+    }
+    // If any remaining descriptors are unused, just point them at src[0]
+    for (uint32_t i = num_tensors; i < MAX_PARAMETER_COUNT; ++i) {
+        buf[i] = buf[0];
+        offset[i] = 0;
+    }
+    if (ctx->do_add_rms_partials) {
+        buf[num_tensors] = ctx->prealloc_add_rms_partials;
+        offset[num_tensors] = ctx->prealloc_size_add_rms_partials_offset;
+    }
+
+    std::array<uint32_t, 3> elements;
+
+    uint32_t ne = ggml_nelements(dst);
+    if (ne > 262144) {
+        elements = { 512, 512, CEIL_DIV(ne, 262144) };
+    } else if (ne > 512) {
+        elements = { 512, CEIL_DIV(ne, 512), 1 };
+    } else {
+        elements = { ne, 1, 1 };
+    }
+
+    static_assert(MAX_PARAMETER_COUNT == 12);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+        {
+            vk_subbuffer{ buf[0], offset[0], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[1], offset[1], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[2], offset[2], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[3], offset[3], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[4], offset[4], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[5], offset[5], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[6], offset[6], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[7], offset[7], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[8], offset[8], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[9], offset[9], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[10], offset[10], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[11], offset[11], VK_WHOLE_SIZE },
+        }, pc, elements);
+}
+
 static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
@@ -7750,7 +8158,7 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        0.0f, 0.0f, 0,
+        0.0f, 0.0f, ctx->do_add_rms_partials,
     }, dryrun);
 }
 
@@ -7838,8 +8246,6 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
         src_buf_ctxs[i] = (ggml_backend_vk_buffer_context *)dst->src[i]->buffer->context;
     }
 
-    ggml_vk_sync_buffers(subctx);
-
     vk_buffer d_D = nullptr, d_srcs[7] = { nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr };
     size_t dst_offset = 0, src_offsets[7] = { 0, 0, 0, 0, 0, 0, 0 };
     bool dst_uma = false, srcs_uma[7] = { false, false, false, false, false, false, false };
@@ -7977,8 +8383,6 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
     ggml_backend_vk_buffer_context * gv_buf_ctx = (ggml_backend_vk_buffer_context *)gv->buffer->context;
     ggml_backend_vk_buffer_context * p_buf_ctx = (ggml_backend_vk_buffer_context *)p->buffer->context;
 
-    ggml_vk_sync_buffers(subctx);
-
     vk_buffer d_X = nullptr, d_G = nullptr, d_GM = nullptr, d_GV = nullptr, d_P = nullptr;
     size_t x_offset = 0, g_offset = 0, gm_offset = 0, gv_offset = 0, p_offset = 0;
     bool X_uma = false, G_uma = false, GM_uma = false, GV_uma = false, P_uma = false;
@@ -8045,6 +8449,12 @@ static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& su
     );
 }
 
+static void ggml_vk_opt_step_sgd(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+    const size_t n = ggml_nelements(dst->src[0]);
+
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun);
+}
+
 static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     int * op_params = (int *)dst->op_params;
 
@@ -8097,6 +8507,10 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
 
+static void ggml_vk_sqrt(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQRT, vk_op_unary_push_constants_init(src0, dst), dryrun);
+}
+
 static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, vk_op_unary_push_constants_init(src0, dst), dryrun);
 }
@@ -8202,19 +8616,39 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
 }
 
+static uint32_t ggml_vk_rms_num_partials(ggml_backend_vk_context * ctx, const ggml_tensor *node) {
+    const uint32_t ne = (uint32_t)node->ne[0];
+    const uint32_t denom = ctx->device->pipeline_add_rms[0][0][0]->wg_denoms[0];
+    const uint32_t num_partials = CEIL_DIV(ne, denom);
+    return num_partials;
+}
+
+static uint32_t ggml_vk_rms_partials_size(ggml_backend_vk_context * ctx, const ggml_tensor *node) {
+    const uint32_t num_partials = ggml_vk_rms_num_partials(ctx, node);
+    const uint32_t num_bytes = ROUNDUP_POW2(num_partials * sizeof(uint32_t), ctx->device->partials_binding_alignment);
+    return num_bytes;
+}
+
 static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, float * op_params, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
+    uint32_t param3 = ctx->do_add_rms_partials ? ggml_vk_rms_num_partials(ctx, dst) : 0;
+
     ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        op_params[0], 0.0f, 0,
+        op_params[0], 0.0f, (int32_t)param3,
     }, dryrun);
+
+    if (ctx->do_add_rms_partials) {
+        ctx->prealloc_size_add_rms_partials_offset += ggml_vk_rms_partials_size(ctx, src0);
+        ctx->do_add_rms_partials = false;
+    }
 }
 
 static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -8336,7 +8770,7 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
         (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
         freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
         src2 != nullptr, (uint32_t)src0->ne[2], s1, s2,
-        sections[0], sections[1], sections[2], sections[3], backprop
+        { sections[0], sections[1], sections[2], sections[3] }, backprop
     }, dryrun);
 }
 
@@ -8345,30 +8779,30 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c
 
     uint32_t ncols = src0->ne[0];
 
-    uint32_t ncols_pad = 1;
-    while (ncols_pad < ncols) {
-        ncols_pad *= 2;
-    }
-
-    GGML_ASSERT(ncols_pad <= 1024);
-
     ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
         ncols,
-        ncols_pad,
         op_params[0],
     }, dryrun);
 }
 
 static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+    vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, ggml_nelements(src0));
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM, p, dryrun);
 }
 
 static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun);
+    vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, p, dryrun);
+}
+
+static void ggml_vk_mean(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    vk_op_sum_rows_push_constants p = vk_op_sum_rows_push_constants_init(src0, dst, src0->ne[0]);
+    p.weight = 1.0f / (float)src0->ne[0];
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_MEAN, p, dryrun);
 }
 
 static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }, dryrun);
 }
 
 static void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -9492,6 +9926,14 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
         }
         ctx->prealloc_split_k = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_split_k);
     }
+    if (ctx->prealloc_add_rms_partials == nullptr || (ctx->prealloc_size_add_rms_partials > 0 && ctx->prealloc_add_rms_partials->size < ctx->prealloc_size_add_rms_partials)) {
+        VK_LOG_MEMORY("ggml_vk_preallocate_buffers(add_partials_size: " << ctx->prealloc_add_rms_partials << ")");
+        // Resize buffer
+        if (ctx->prealloc_add_rms_partials != nullptr) {
+            ggml_vk_destroy_buffer(ctx->prealloc_add_rms_partials);
+        }
+        ctx->prealloc_add_rms_partials = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_add_rms_partials);
+    }
 }
 
 static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, ggml_tensor* tensor, int tensor_idx, bool use_fence, bool almost_ready);
@@ -9522,6 +9964,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         return false;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
+        case GGML_UNARY_OP_EXP:
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
         case GGML_UNARY_OP_GELU_ERF:
@@ -9547,10 +9990,23 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
             return false;
         }
         break;
+    case GGML_OP_ADD:
+        {
+            int next_node_idx = node_idx + 1 + ctx->num_additional_fused_ops;
+            if (next_node_idx < cgraph->n_nodes &&
+                cgraph->nodes[next_node_idx]->op == GGML_OP_RMS_NORM &&
+                cgraph->nodes[next_node_idx]->src[0] == cgraph->nodes[next_node_idx - 1] &&
+                ggml_nrows(cgraph->nodes[next_node_idx]) == 1 &&
+                ctx->device->add_rms_fusion) {
+                if (dryrun) {
+                    ctx->prealloc_size_add_rms_partials += ggml_vk_rms_partials_size(ctx, cgraph->nodes[node_idx]);
+                }
+                ctx->do_add_rms_partials = true;
+            }
+        } break;
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_GET_ROWS:
-    case GGML_OP_ADD:
     case GGML_OP_ADD_ID:
     case GGML_OP_ACC:
     case GGML_OP_SUB:
@@ -9560,6 +10016,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
@@ -9585,6 +10042,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_ARGSORT:
     case GGML_OP_SUM:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
     case GGML_OP_ARGMAX:
     case GGML_OP_COUNT_EQUAL:
     case GGML_OP_IM2COL:
@@ -9598,11 +10056,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_FLASH_ATTN_EXT:
     case GGML_OP_OPT_STEP_ADAMW:
+    case GGML_OP_OPT_STEP_SGD:
         break;
     default:
         std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
         GGML_ABORT("fatal error");
-        return false;
     }
 
     vk_context compute_ctx;
@@ -9629,6 +10087,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_UPSCALE:
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
+        case GGML_OP_SQRT:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
@@ -9653,6 +10112,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_ARGSORT:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_ARGMAX:
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_IM2COL:
@@ -9662,11 +10122,15 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_CONV_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_OPT_STEP_SGD:
             {
                 // These operations all go through ggml_vk_op_f32, so short-circuit and
                 // do the only thing needed for the dryrun.
                 vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op);
                 ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+                if (node->op == GGML_OP_RMS_NORM) {
+                    ctx->do_add_rms_partials = false;
+                }
                 return false;
             }
         default:
@@ -9674,6 +10138,83 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         }
     }
 
+    if (!dryrun) {
+        // This logic detects dependencies between modes in the graph and calls ggml_vk_sync_buffers
+        // to synchronize them. This handles most "normal" synchronization when computing the graph, and when
+        // there is no auxiliary memory use, it shouldn't be necessary to call ggml_vk_sync_buffers
+        // outside of this logic. When a node uses one of the prealloc buffers for something like
+        // dequantization or split_k, additional synchronization is needed between those passes.
+        bool need_sync = false;
+
+        // Check whether "node" requires synchronization. The node requires synchronization if it
+        // overlaps in memory with another unsynchronized node and at least one of them is a write.
+        // Destination nodes are checked against both the written/read lists. Source nodes are only
+        // checked against the written list. Two nodes overlap in memory if they come from the same
+        // buffer and the tensor or view ranges overlap.
+        auto const &overlaps_unsynced = [&](const ggml_tensor *node, const std::vector<const ggml_tensor *> &unsynced_nodes) -> bool {
+            if (unsynced_nodes.size() == 0) {
+                return false;
+            }
+            auto n_base = vk_tensor_offset(node) + node->view_offs;
+            auto n_size = ggml_nbytes(node);
+            ggml_backend_vk_buffer_context * a_buf_ctx = (ggml_backend_vk_buffer_context *)node->buffer->context;
+            vk_buffer a_buf = a_buf_ctx->dev_buffer;
+            for (auto &other : unsynced_nodes) {
+                ggml_backend_vk_buffer_context * o_buf_ctx = (ggml_backend_vk_buffer_context *)other->buffer->context;
+                vk_buffer o_buf = o_buf_ctx->dev_buffer;
+                if (a_buf == o_buf) {
+                    auto o_base = vk_tensor_offset(other) + other->view_offs;
+                    auto o_size = ggml_nbytes(other);
+
+                    if ((o_base <= n_base && n_base < o_base + o_size) ||
+                        (n_base <= o_base && o_base < n_base + n_size)) {
+                        return true;
+                    }
+                }
+            }
+            return false;
+        };
+
+        // For all fused ops, check if the destination node or any of the source
+        // nodes require synchronization.
+        for (int32_t i = 0; i < ctx->num_additional_fused_ops + 1 && !need_sync; ++i) {
+            const ggml_tensor *cur_node = cgraph->nodes[node_idx + i];
+            if (overlaps_unsynced(cur_node, ctx->unsynced_nodes_read) || overlaps_unsynced(cur_node, ctx->unsynced_nodes_written)) {
+                need_sync = true;
+                break;
+            }
+            for (uint32_t j = 0; j < GGML_MAX_SRC; ++j) {
+                if (!cur_node->src[j]) {
+                    continue;
+                }
+                if (overlaps_unsynced(cur_node->src[j], ctx->unsynced_nodes_written)) {
+                    need_sync = true;
+                    break;
+                }
+            }
+        }
+        if (need_sync) {
+            VK_LOG_DEBUG("node_idx=" << i << " sync");
+            ctx->unsynced_nodes_written.clear();
+            ctx->unsynced_nodes_read.clear();
+            ggml_vk_sync_buffers(ctx, compute_ctx);
+        } else {
+            VK_LOG_DEBUG("node_idx=" << i << " unsynced");
+        }
+        // Add the last fused node and all fused source nodes to the unsynchronized list.
+        const ggml_tensor * last_node = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
+        ctx->unsynced_nodes_written.push_back(last_node);
+        for (int32_t i = 0; i < ctx->num_additional_fused_ops + 1; ++i) {
+            const ggml_tensor *cur_node = cgraph->nodes[node_idx + i];
+            for (uint32_t j = 0; j < GGML_MAX_SRC; ++j) {
+                if (!cur_node->src[j]) {
+                    continue;
+                }
+                ctx->unsynced_nodes_read.push_back(cur_node->src[j]);
+            }
+        }
+    }
+
     switch (node->op) {
     case GGML_OP_REPEAT:
         ggml_vk_repeat(ctx, compute_ctx, src0, node, dryrun);
@@ -9692,8 +10233,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         break;
     case GGML_OP_ADD:
-        ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
-
+        if (ctx->num_additional_fused_ops) {
+            ggml_vk_multi_add(ctx, compute_ctx, cgraph, node_idx, dryrun);
+        } else {
+            ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
+        }
         break;
     case GGML_OP_SUB:
         ggml_vk_sub(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -9726,6 +10270,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_SQR:
         ggml_vk_sqr(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_SQRT:
+        ggml_vk_sqrt(ctx, compute_ctx, src0, node, dryrun);
+
         break;
     case GGML_OP_SIN:
         ggml_vk_sin(ctx, compute_ctx, src0, node, dryrun);
@@ -9789,6 +10337,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(node)) {
+        case GGML_UNARY_OP_EXP:
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
         case GGML_UNARY_OP_GELU_ERF:
@@ -9847,6 +10396,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_SUM_ROWS:
         ggml_vk_sum_rows(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_MEAN:
+        ggml_vk_mean(ctx, compute_ctx, src0, node, dryrun);
+
         break;
     case GGML_OP_ARGMAX:
         ggml_vk_argmax(ctx, compute_ctx, src0, node, dryrun);
@@ -9911,6 +10464,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_OPT_STEP_ADAMW:
         ggml_vk_opt_step_adamw(ctx, compute_ctx, node, dryrun);
 
+        break;
+
+    case GGML_OP_OPT_STEP_SGD:
+        ggml_vk_opt_step_sgd(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+
         break;
     default:
         return false;
@@ -9972,6 +10530,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_UPSCALE:
     case GGML_OP_SCALE:
     case GGML_OP_SQR:
+    case GGML_OP_SQRT:
     case GGML_OP_SIN:
     case GGML_OP_COS:
     case GGML_OP_CLAMP:
@@ -10000,6 +10559,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_ARGSORT:
     case GGML_OP_SUM:
     case GGML_OP_SUM_ROWS:
+    case GGML_OP_MEAN:
     case GGML_OP_ARGMAX:
     case GGML_OP_COUNT_EQUAL:
     case GGML_OP_IM2COL:
@@ -10014,11 +10574,12 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_OPT_STEP_ADAMW:
+    case GGML_OP_OPT_STEP_SGD:
         buf = tensor->buffer;
-
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(tensor)) {
+        case GGML_UNARY_OP_EXP:
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
         case GGML_UNARY_OP_GELU_ERF:
@@ -10115,6 +10676,11 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
         ggml_vk_pool_free(ctx, buffer);
     }
     ctx->gc.temp_buffers.clear();
+    ctx->prealloc_y_last_pipeline_used = {};
+
+    ctx->unsynced_nodes_written.clear();
+    ctx->unsynced_nodes_read.clear();
+    ctx->prealloc_x_need_sync = ctx->prealloc_y_need_sync = ctx->prealloc_split_k_need_sync = false;
 
     ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
     ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
@@ -10150,6 +10716,7 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     ggml_vk_destroy_buffer(ctx->prealloc_x);
     ggml_vk_destroy_buffer(ctx->prealloc_y);
     ggml_vk_destroy_buffer(ctx->prealloc_split_k);
+    ctx->prealloc_y_last_pipeline_used = nullptr;
 
     for (auto& buffer : ctx->buffer_pool) {
         ggml_vk_destroy_buffer(buffer);
@@ -10570,6 +11137,58 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
     return true;
 }
 
+static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
+
+    const ggml_tensor *first_node = cgraph->nodes[node_idx];
+    if (first_node->op != GGML_OP_ADD) {
+        return 0;
+    }
+
+    if (!ctx->device->multi_add) {
+        return 0;
+    }
+
+    int32_t num_adds = 1;
+    while (node_idx + num_adds < cgraph->n_nodes &&
+           cgraph->nodes[node_idx + num_adds]->op == GGML_OP_ADD &&
+           num_adds < MAX_FUSED_ADDS) {
+        num_adds++;
+    }
+
+    // The shader currently requires same shapes (but different strides are allowed),
+    // everything f32, and no misalignment
+    for (int32_t i = 0; i < num_adds; ++i) {
+        const ggml_tensor *next_node = cgraph->nodes[node_idx + i];
+        if (!ggml_are_same_shape(first_node, next_node->src[0]) ||
+            !ggml_are_same_shape(first_node, next_node->src[1]) ||
+            next_node->type != GGML_TYPE_F32 ||
+            next_node->src[0]->type != GGML_TYPE_F32 ||
+            next_node->src[1]->type != GGML_TYPE_F32 ||
+            get_misalign_bytes(ctx, next_node) ||
+            get_misalign_bytes(ctx, next_node->src[0]) ||
+            get_misalign_bytes(ctx, next_node->src[1])) {
+            num_adds = i;
+        }
+    }
+
+    // Verify we can fuse these
+    ggml_op adds[MAX_FUSED_ADDS];
+    for (int32_t i = 0; i < num_adds; ++i) {
+        adds[i] = GGML_OP_ADD;
+    }
+
+    // decrease num_adds if they can't all be fused
+    while (num_adds > 1 && !ggml_can_fuse(cgraph, node_idx, adds, num_adds)) {
+        num_adds--;
+    }
+
+    // a single add is not "fused", so just return zero
+    if (num_adds == 1) {
+        return 0;
+    }
+    return num_adds;
+}
+
 static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
@@ -10581,10 +11200,19 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue.queue, reinterpret_cast<VkDebugUtilsLabelEXT*>(&dul));
     }
 
+    ctx->prealloc_size_add_rms_partials = 0;
+    ctx->prealloc_size_add_rms_partials_offset = 0;
+    ctx->do_add_rms_partials = false;
+
     uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-            ctx->num_additional_fused_ops = 1;
+        if (!ctx->device->disable_fusion) {
+            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
+            if (num_adds) {
+                ctx->num_additional_fused_ops = num_adds - 1;
+            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            }
         }
         ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
         if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
@@ -10641,6 +11269,22 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         compute_ctx->s->buffer.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->device->query_pool, 0);
     }
 
+    ctx->prealloc_y_last_pipeline_used = nullptr;
+    ctx->prealloc_y_last_tensor_used = nullptr;
+
+    if (ctx->prealloc_size_add_rms_partials) {
+        if (ctx->compute_ctx.expired()) {
+            compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+            ctx->compute_ctx = compute_ctx;
+            ggml_vk_ctx_begin(ctx->device, compute_ctx);
+        } else {
+            compute_ctx = ctx->compute_ctx.lock();
+        }
+        // initialize partial sums to zero.
+        ggml_vk_buffer_memset_async(compute_ctx, ctx->prealloc_add_rms_partials, 0, 0, ctx->prealloc_size_add_rms_partials);
+        ggml_vk_sync_buffers(ctx, compute_ctx);
+    }
+
     // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
     // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB
     // (and scaled down based on model size, so smaller models submit earlier).
@@ -10659,8 +11303,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
         }
 
-        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-            ctx->num_additional_fused_ops = 1;
+        if (!ctx->device->disable_fusion) {
+            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
+            if (num_adds) {
+                ctx->num_additional_fused_ops = num_adds - 1;
+            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            }
         }
 
         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
@@ -10867,6 +11516,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_EXP:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_GELU_QUICK:
@@ -10881,7 +11531,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 default:
                     return false;
             }
-            break;
         case GGML_OP_GLU:
             switch (ggml_get_glu_op(op)) {
                 case GGML_GLU_OP_GEGLU:
@@ -10897,7 +11546,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 default:
                     return false;
             }
-            break;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
@@ -10961,14 +11609,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 }
 
                 return true;
-            } break;
+            }
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
                 auto device = ggml_vk_get_device(ctx->device);
                 bool coopmat2 = device->coopmat2;
-                FaHeadSizes head_sizes = fa_get_head_sizes(op->src[1]->ne[0], op->src[2]->ne[0]);
-                if (head_sizes == FA_HEAD_SIZE_UNSUPPORTED) {
+                uint32_t HSK = op->src[1]->ne[0];
+                uint32_t HSV = op->src[2]->ne[0];
+                if ((HSK % 8) != 0 || (HSV % 8) != 0) {
                     return false;
                 }
                 if (op->src[4] && op->src[4]->type != GGML_TYPE_F32) {
@@ -11051,7 +11700,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     default:
                         return false;
                 }
-            } break;
+            }
         case GGML_OP_SET_ROWS:
             {
                 switch (op->type) {
@@ -11068,7 +11717,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     default:
                         return false;
                 }
-            } break;
+            }
         case GGML_OP_CONT:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
@@ -11120,7 +11769,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     return true;
                 }
                 return false;
-            } break;
+            }
         case GGML_OP_REPEAT:
             return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float);
         case GGML_OP_REPEAT_BACK:
@@ -11151,10 +11800,16 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_SQR:
+        case GGML_OP_SQRT:
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+        case GGML_OP_LEAKY_RELU:
+        case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_ARGSORT:
+            return op->ne[0] <= max_argsort_cols;
         case GGML_OP_UPSCALE:
         case GGML_OP_ACC:
         case GGML_OP_CONCAT:
@@ -11164,9 +11819,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
-        case GGML_OP_ARGSORT:
+            return true;
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
+            return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_ARGMAX:
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_IM2COL:
@@ -11175,8 +11832,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_POOL_2D:
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
-        case GGML_OP_LEAKY_RELU:
-        case GGML_OP_OPT_STEP_ADAMW:
             return true;
         case GGML_OP_CONV_TRANSPOSE_1D:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
@@ -11185,14 +11840,13 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 // Op is disabled for Apple because it segfaults at pipeline create time on MoltenVK
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
                 const vk_device& device = ggml_vk_get_device(ctx->device);
-                bool is_Apple = ggml_vk_get_device(ctx->device)->vendor_id == VK_VENDOR_ID_APPLE;
                 // Channel-contiguous format is not supported yet.
                 return ((op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
                     op->src[1]->type == GGML_TYPE_F32 &&
                     op->type == GGML_TYPE_F32 &&
                     ggml_is_contiguous(op->src[0]) &&
                     ggml_is_contiguous(op->src[1]) &&
-                    ggml_is_contiguous(op)) && !is_Apple;
+                    ggml_is_contiguous(op));
             }
         default:
             return false;
@@ -11593,6 +12247,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
     } else if (tensor->op == GGML_OP_SQR) {
         tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
+    } else if (tensor->op == GGML_OP_SQRT) {
+        tensor_clone = ggml_sqrt(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_SIN) {
         tensor_clone = ggml_sin(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_COS) {
@@ -11663,6 +12319,9 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         }
     } else if (tensor->op == GGML_OP_UNARY) {
         switch (ggml_get_unary_op(tensor)) {
+        case GGML_UNARY_OP_EXP:
+            tensor_clone = ggml_exp(ggml_ctx, src_clone[0]);
+            break;
         case GGML_UNARY_OP_SILU:
             tensor_clone = ggml_silu(ggml_ctx, src_clone[0]);
             break;
@@ -11694,6 +12353,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else {
             tensor_clone = ggml_glu_split(ggml_ctx, src_clone[0], src_clone[1], (ggml_glu_op) tensor->op_params[0]);
         }
+        ggml_set_op_params_i32(tensor_clone, 2, ggml_get_op_params_i32(tensor, 2));
+        ggml_set_op_params_i32(tensor_clone, 3, ggml_get_op_params_i32(tensor, 3));
     } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
         if (src1 == nullptr) {
             tensor_clone = ggml_dup(ggml_ctx, src_clone[0]);
@@ -11720,6 +12381,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         tensor_clone = ggml_sum(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_SUM_ROWS) {
         tensor_clone = ggml_sum_rows(ggml_ctx, src_clone[0]);
+    } else if (tensor->op == GGML_OP_MEAN) {
+        tensor_clone = ggml_mean(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_ARGMAX) {
         tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_COUNT_EQUAL) {
@@ -11774,6 +12437,12 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         src_clone[0]->flags = src0->flags;
         tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1],
         src_clone[2], src_clone[3], src_clone[4]);
+    } else if (tensor->op == GGML_OP_OPT_STEP_SGD) {
+        src_clone[0]->flags = src0->flags;
+        tensor_clone = ggml_opt_step_sgd(ggml_ctx, src_clone[0], src_clone[1],
+        src_clone[2]);
+    } else if (tensor->op == GGML_OP_ADD_ID) {
+        tensor_clone = ggml_add_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
     }
     else {
         std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/add.comp b/ggml/src/ggml-vulkan/vulkan-shaders/add.comp
index 2b4085c4f82d5..00cf2dd62fddb 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/add.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/add.comp
@@ -1,20 +1,34 @@
 #version 450
 
 #extension GL_EXT_shader_16bit_storage : require
+#if ADD_RMS
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#endif
 
 #include "types.comp"
 #include "generic_binary_head.comp"
 
 const uint num_threads = 256;
 
+layout (binding = 3, std430) buffer PartialBuf {float partial_sums[];};
+
 layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
 
+#if ADD_RMS
+// XXX TODO this could be sized based on number of subgroups, but that't not considered a constant
+shared FLOAT_TYPE sumsh[num_threads];
+#endif
+
 void main() {
     uint idx = get_idx();
+    uint orig_idx = idx;
 
     // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
     const uint num_iter = 2;
 
+    FLOAT_TYPE sum_sq = 0;
+
     [[unroll]] for (uint i = 0; i < num_iter; ++i) {
         if (idx >= p.ne) {
             continue;
@@ -22,8 +36,34 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
+        FLOAT_TYPE sum = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]);
+        sum_sq += sum*sum;
+
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
 
         idx += num_threads;
     }
+
+#if ADD_RMS
+    if (p.param3 != 0) {
+        // reduce the sum within each subgroup, then across subgroups
+        const uint NumSubgroups = num_threads / gl_SubgroupSize;
+        sum_sq = subgroupAdd(sum_sq);
+        if (gl_SubgroupInvocationID == 0) {
+            sumsh[gl_SubgroupID] = sum_sq;
+        }
+        barrier();
+        [[unroll]] for (uint s = NumSubgroups / 2; s > 0; s >>= 1) {
+            if (gl_SubgroupID < s && gl_SubgroupInvocationID == 0) {
+                sum_sq += sumsh[gl_SubgroupID + s];
+                sumsh[gl_SubgroupID] = sum_sq;
+            }
+            barrier();
+        }
+
+        if (gl_SubgroupID == 0 && gl_SubgroupInvocationID == 0) {
+            partial_sums[orig_idx / (num_iter * num_threads)] = sum_sq;
+        }
+    }
+#endif
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp b/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp
index eaf4da341e348..a1d4c240ddf71 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp
@@ -5,6 +5,8 @@
 
 #extension GL_EXT_control_flow_attributes : enable
 
+#define FLT_MAX 3.402823466e+38F
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -19,19 +21,26 @@ void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
 
-    if (col >= p.KX) {
+    if (row >= p.KY) {
         return;
     }
-    A_TYPE amax = data_a[row*p.KX + col];
-    tmp[col] = col;
+
+    A_TYPE amax = -FLT_MAX;
+    uint acol = col;
+
+    if (col < p.KX) {
+        amax = data_a[row*p.KX + col];
+    }
 
     for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) {
         A_TYPE val = data_a[row*p.KX + i];
         if (val > amax) {
             amax = val;
-            tmp[col] = i;
+            acol = i;
         }
     }
+
+    tmp[col] = acol;
     tmpmax[col] = amax;
 
     barrier();
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp b/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
index d4fa45b1e106f..dc53a401e0ba7 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
@@ -1,22 +1,24 @@
 #version 450
+#extension GL_EXT_control_flow_attributes : enable
 
 #include "types.comp"
 
-#define BLOCK_SIZE 1024
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
 #define ASC 0
 
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1)          buffer D {int data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
-    uint ncols_pad;
     uint order;
 } p;
 
 shared int dst_row[BLOCK_SIZE];
+shared A_TYPE a_sh[BLOCK_SIZE];
 
 void swap(uint idx0, uint idx1) {
     int tmp = dst_row[idx0];
@@ -24,7 +26,7 @@ void swap(uint idx0, uint idx1) {
     dst_row[idx1] = tmp;
 }
 
-void main() {
+void argsort(bool needs_bounds_check) {
     // bitonic sort
     const int col = int(gl_LocalInvocationID.x);
     const uint row = gl_WorkGroupID.y;
@@ -32,38 +34,46 @@ void main() {
     const uint row_offset = row * p.ncols;
 
     // initialize indices
-    if (col < p.ncols_pad) {
-        dst_row[col] = col;
-    }
+    dst_row[col] = col;
+    a_sh[col] = data_a[row_offset + col];
     barrier();
 
-    for (uint k = 2; k <= p.ncols_pad; k *= 2) {
-        for (uint j = k / 2; j > 0; j /= 2) {
-            const uint ixj = col ^ j;
-            if (col < p.ncols_pad && ixj > col) {
-                if ((col & k) == 0) {
-                    if (dst_row[col] >= p.ncols ||
-                        (dst_row[ixj] < p.ncols && (p.order == ASC ?
-                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
-                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
-                    ) {
-                        swap(col, ixj);
-                    }
-                } else {
-                    if (dst_row[ixj] >= p.ncols ||
-                        (dst_row[col] < p.ncols && (p.order == ASC ?
-                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
-                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
-                    ) {
-                        swap(col, ixj);
-                    }
-                }
+    uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
+    [[unroll]] for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
+        uint num_inner_loop_iters = outer_idx + 1;
+        [[unroll]] for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
+            const int ixj = int(col ^ j);
+
+            int idx_0 = (col & k) == 0 ? col : ixj;
+            int idx_1 = (col & k) == 0 ? ixj : col;
+
+            int sh_idx_0 = dst_row[idx_0];
+            int sh_idx_1 = dst_row[idx_1];
+            bool idx_0_oob = needs_bounds_check ? sh_idx_0 >= p.ncols : false;
+            bool idx_1_oob = needs_bounds_check ? sh_idx_1 >= p.ncols : false;
+
+            if ((idx_0_oob ||
+                (!idx_1_oob && a_sh[sh_idx_0] > a_sh[sh_idx_1])) && (ixj > col)) {
+                swap(idx_0, idx_1);
             }
+
             barrier();
         }
     }
 
     if (col < p.ncols) {
-        data_d[row_offset + col] = dst_row[col];
+        if (p.order == ASC) {
+            data_d[row_offset + col] = dst_row[col];
+        } else {
+            data_d[row_offset + p.ncols - col - 1] = dst_row[col];
+        }
+    }
+}
+
+void main() {
+    if (p.ncols == BLOCK_SIZE) {
+        argsort(false);
+    } else {
+        argsort(true);
     }
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp b/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp
new file mode 100644
index 0000000000000..abecd2d3dcf74
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp
@@ -0,0 +1,20 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+    data_d[i] = D_TYPE(exp(float(data_a[i])));
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
index b57c9dcfc4ee5..f73e17e1fa8d9 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
@@ -9,6 +9,10 @@ layout (constant_id = 4) const uint32_t HSV = 32;
 layout (constant_id = 5) const uint32_t Clamp = 0;
 layout (constant_id = 6) const uint32_t D_split = 16;
 
+// Round up head sizes to a multiple of 16, for coopmat1/coopmat2 paths
+const uint32_t HSK_pad = (HSK + 15) & ~15;
+const uint32_t HSV_pad = (HSV + 15) & ~15;
+
 layout (push_constant) uniform parameter {
     uint32_t N;
     uint32_t KV;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
index 230e815f22c45..97c2a54129709 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
@@ -46,14 +46,14 @@ const uint32_t MatBc = 16;
 shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
 shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
 
-const uint32_t qstride = HSK / 4 + 2; // in units of f16vec4
+const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
 // Avoid padding for hsk==256 to make it fit in 48KB shmem.
 const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br;
 shared ACC_TYPE sfsh[Bc * sfshstride];
 
-const uint32_t kshstride = HSK / 4 + 2; // in units of f16vec4
+const uint32_t kshstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 ksh[Bc * kshstride];
 
 shared float slope[Br];
@@ -74,6 +74,21 @@ void main() {
 
 #define tile_row(r) (row_tid * rows_per_thread + (r))
 
+    // Zero-initialize shared memory for Q/K when HSK is not a multiple of 16 (HSK_pad > HSK).
+    if ((HSK % 16) != 0) {
+        [[unroll]] for (uint i = 0; i < Br * qstride; i += gl_WorkGroupSize.x) {
+            if (i + tid < Br * qstride) {
+                Qf[i + tid] = f16vec4(0);
+            }
+        }
+        [[unroll]] for (uint i = 0; i < Bc * kshstride; i += gl_WorkGroupSize.x) {
+            if (i + tid < Bc * kshstride) {
+                ksh[i + tid] = f16vec4(0);
+            }
+        }
+        barrier();
+    }
+
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
     [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
@@ -151,14 +166,14 @@ void main() {
         }
         barrier();
 
-        // K * Q^T -> S^T: Bc x HSK * HSK x Br -> Bc x Br
+        // K * Q^T -> S^T: Bc x HSK_pad * HSK_pad x Br -> Bc x Br
         // Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
         // This is written transposed in order to allow for N being 8 if implementations need it
         coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
         coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
         coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
 
-        for (uint32_t d = 0; d < HSK / 16; ++d) {
+        for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
             coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
 
             uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
@@ -210,7 +225,7 @@ void main() {
 
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+                Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             }
         }
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -233,7 +248,7 @@ void main() {
                 vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
 #endif
                 [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                    Of[r][d] += float16_t(Pf[r]) * ACC_TYPEV4(Vf);
+                    Of[r][d] += ACC_TYPE(Pf[r]) * ACC_TYPEV4(Vf);
                 }
             }
         }
@@ -288,7 +303,7 @@ void main() {
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
-            Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+            Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
 
             barrier();
@@ -357,7 +372,7 @@ void main() {
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            Of[r][d] *= float16_t(Lfrcp[r]);
+            Of[r][d] *= ACC_TYPE(Lfrcp[r]);
         }
     }
 
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
index b0564ca0bfc83..77ae5ff01d03e 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -104,16 +104,16 @@ void main() {
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
     tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
 
-    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseAccumulator> Q;
-    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA> Qf16;
+    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseAccumulator> Q;
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA> Qf16;
 
     uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
-    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK));
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK_pad));
 
-    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA>(Q);
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -140,10 +140,10 @@ void main() {
 
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, HSK, Bc, gl_MatrixUseB> K_T;
+        coopmat<float16_t, gl_ScopeWorkgroup, HSK_pad, Bc, gl_MatrixUseB> K_T;
 
         uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
-        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK), tensorViewTranspose DECODEFUNC);
+        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK_pad), tensorViewTranspose DECODEFUNC);
         S = coopMatMulAdd(Qf16, K_T, S);
 
         if (p.logit_softcap != 0.0f) {
@@ -208,31 +208,31 @@ void main() {
         rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
         rowsum = coopMatMulAdd(P_A, One, rowsum);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV, gl_MatrixUseB> V;
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV_pad, gl_MatrixUseB> V;
         uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
-        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV) DECODEFUNC);
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV_pad) DECODEFUNC);
 
         L = eM*L + rowsum;
 
         // This is the "diagonal" matrix in the paper, but since we do componentwise
         // multiply rather than matrix multiply it has the diagonal element smeared
         // across the row
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> eMdiag;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> eMdiag;
 
         // resize eM by using smear/reduce
         coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
         // multiply with fp16 accumulation, then add to O.
-        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
         PV = coopMatMulAdd(P_A, V, PV);
 
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(PV);
+        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(PV);
     }
 
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
+        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
 
         uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
@@ -243,16 +243,16 @@ void main() {
         return;
     }
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Ldiag;
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Ldiag;
 
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
     if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> S;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> S;
         coopMatPerElementNV(S, S, perElemOpGetSink, iq2);
 
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Mr;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Mr;
 
         // resize M by using smear/reduce
         coopMatReduceNV(Mr, M, gl_CooperativeMatrixReduceRowNV, smearReduce);
@@ -285,7 +285,7 @@ void main() {
 
     uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
-    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
     } else {
@@ -295,6 +295,6 @@ void main() {
         // permute dimensions
         tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
 
-        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV), tensorViewPermute);
+        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV_pad), tensorViewPermute);
     }
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
index 4b4316cf3d9f2..750e785753954 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
@@ -2,6 +2,7 @@
 #extension GL_EXT_control_flow_attributes : require
 
 #include "rte.comp"
+#include "utils.comp"
 
 layout (push_constant) uniform parameter
 {
@@ -28,25 +29,9 @@ uint get_aoffset() { return p.misalign_offsets >> 16; }
 uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
 uint get_doffset() { return p.misalign_offsets & 0xFF; }
 
-// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
-uint fastmod(uint a, uint b) {
-    if ((b & (b-1)) == 0) {
-        return a & (b-1);
-    }
-    return a % b;
-}
-
-uint fastdiv(uint a, uint b) {
-    return (a < b) ? 0 : (a / b);
-}
 
 void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
-    i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    get_indices(idx, i00, i01, i02, i03, p.ne00, p.ne01, p.ne02, p.ne03);
 }
 
 uint src0_idx(uint i00, uint i01, uint i02, uint i03) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
index 903753c7e2ec5..b93e9948f7fc9 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
@@ -1,6 +1,10 @@
 #extension GL_EXT_control_flow_attributes : enable
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
+#if USE_SUBGROUP_ADD
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#endif
 
 #ifdef MUL_MAT_ID
 #define EXPERT_COUNT 8
@@ -90,7 +94,38 @@ layout (constant_id = 2) const uint NUM_COLS = 1;
 
 shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
 
-void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    // subgroupAdd is probably faster on devices that support it,
+    // particularly when the workgroup has more than one subgroup
+#if USE_SUBGROUP_ADD
+    // sum up partial sums within a subgroup
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            temp[j][n] = subgroupAdd(temp[j][n]);
+        }
+    }
+
+    // Go through shared memory to sum partials across subgroups
+    if (gl_SubgroupInvocationID == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[j][n][gl_SubgroupID] = temp[j][n];
+            }
+        }
+    }
+    barrier();
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                temp[j][n] = FLOAT_TYPE(0);
+                [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                    temp[j][n] += tmpsh[j][n][s];
+                }
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
+            }
+        }
+    }
+#else
     // sum up partial sums and write back result
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -115,4 +150,5 @@ void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32
             }
         }
     }
+#endif
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
index 8c5114a79d23c..d57cc6bdec5df 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -103,16 +103,74 @@ layout (constant_id = 10) const uint WARP = 32;
 shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE];
 shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE];
 
+#define NUM_WARPS (BLOCK_SIZE / WARP)
+
 #ifdef MUL_MAT_ID
 shared u16vec2 row_ids[4096];
 uint _ne1;
 #ifdef COOPMAT
-shared uint _ne1_sh;
+shared uvec4 ballots_sh[NUM_WARPS];
+void load_row_ids(uint expert_idx, bool nei0_is_pow2) {
+    _ne1 = 0;
+    uint num_elements = p.nei1 * p.nei0;
+    uint nei0shift = findLSB(p.nei0);
+
+    uint ids[16];
+    uint iter = 0;
+
+    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
+        // prefetch up to 16 elements
+        if (iter == 0) {
+            [[unroll]] for (uint k = 0; k < 16; ++k) {
+                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
+                bool in_range = i < num_elements;
+                uint ii1;
+                if (nei0_is_pow2) {
+                    ii1 = i >> nei0shift;
+                } else {
+                    ii1 = i / p.nei0;
+                }
+                uint ii0 = i - ii1 * p.nei0;
+                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            }
+        }
+        uint i = j + gl_LocalInvocationIndex;
+        bool in_range = i < num_elements;
+        uint ii1;
+        if (nei0_is_pow2) {
+            ii1 = i >> nei0shift;
+        } else {
+            ii1 = i / p.nei0;
+        }
+        uint ii0 = i - ii1 * p.nei0;
+        uint id = ids[iter++];
+        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+
+        ballots_sh[gl_SubgroupID] = ballot;
+        barrier();
+
+        uint subgroup_base = 0;
+        uint total = 0;
+        for (uint k = 0; k < gl_NumSubgroups; ++k) {
+            if (k == gl_SubgroupID) {
+                subgroup_base = total;
+            }
+            total += subgroupBallotBitCount(ballots_sh[k]);
+        }
+        barrier();
+
+        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
+        if (in_range && id == expert_idx) {
+            row_ids[_ne1 + idx] = u16vec2(ii0, ii1);
+        }
+        _ne1 += total;
+        iter &= 15;
+    }
+    barrier();
+}
 #endif
 #endif // MUL_MAT_ID
 
-#define NUM_WARPS (BLOCK_SIZE / WARP)
-
 #ifdef COOPMAT
 shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #endif
@@ -178,44 +236,11 @@ void main() {
 
 #ifdef MUL_MAT_ID
 #ifdef COOPMAT
-    // Spread the search across all elements in the first subgroup
-    if (gl_SubgroupID == 0) {
-        _ne1 = 0;
-        uint num_elements = p.nei1 * p.nei0;
-
-        uint ids[16];
-        uint iter = 0;
-
-        for (uint j = 0; j < num_elements; j += gl_SubgroupSize) {
-            // prefetch up to 16 elements
-            if (iter == 0) {
-                [[unroll]] for (uint k = 0; k < 16; ++k) {
-                    uint i = j + gl_SubgroupInvocationID + k*gl_SubgroupSize;
-                    bool in_range = i < num_elements;
-                    uint ii1 = i / p.nei0;
-                    uint ii0 = i % p.nei0;
-                    ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
-                }
-            }
-            uint i = j + gl_SubgroupInvocationID;
-            bool in_range = i < num_elements;
-            uint ii1 = i / p.nei0;
-            uint ii0 = i % p.nei0;
-            uint id = ids[iter++];
-            uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
-            uint idx = subgroupBallotExclusiveBitCount(ballot);
-            if (in_range && id == expert_idx) {
-                row_ids[_ne1 + idx] = u16vec2(ii0, ii1);
-            }
-            _ne1 += subgroupBallotBitCount(ballot);
-            iter &= 15;
-        }
-        _ne1_sh = _ne1;
+    if (bitCount(p.nei0) == 1) {
+        load_row_ids(expert_idx, true);
+    } else {
+        load_row_ids(expert_idx, false);
     }
-
-    barrier();
-
-    _ne1 = _ne1_sh;
 #else
     _ne1 = 0;
     for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
@@ -801,7 +826,7 @@ void main() {
             }
 #else
             const uint row_i = ic * BN + loadc_b + l;
-            if (row_i < _ne1) {
+            if (row_i < _ne1 && block + loadr_b < end_k) {
                 const u16vec2 row_idx = row_ids[row_i];
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
             } else {
@@ -875,7 +900,9 @@ void main() {
 
                 const u16vec2 row_idx = row_ids[row_i];
 
-                data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                if (dr + cm_row * TM + store_r < p.M) {
+                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                }
             }
         }
     }
@@ -925,7 +952,9 @@ void main() {
 #endif // MUL_MAT_ID
                 [[unroll]] for (uint cr = 0; cr < TM; cr++) {
 #ifdef MUL_MAT_ID
-                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    if (dr_warp + cr < p.M) {
+                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    }
 #else
                     if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
                         data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
index 29e4b5c9ce2d4..4d16eb0791ddc 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
@@ -19,6 +19,7 @@
 #endif
 
 #include "types.comp"
+#include "utils.comp"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
@@ -99,7 +100,8 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB {
 };
 
 uint _ne1;
-shared uint _ne1_sh;
+layout (constant_id = 5) const uint subgroup_size = 32;
+shared uvec4 ballots_sh[BLOCK_SIZE / subgroup_size];
 
 B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
@@ -128,6 +130,64 @@ D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem
     return elem;
 }
 
+void load_row_ids(uint expert_idx, bool nei0_is_pow2) {
+    _ne1 = 0;
+    uint num_elements = p.nei1 * p.nei0;
+    uint nei0shift = findLSB(p.nei0);
+
+    uint ids[16];
+    uint iter = 0;
+
+    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
+        // prefetch up to 16 elements
+        if (iter == 0) {
+            [[unroll]] for (uint k = 0; k < 16; ++k) {
+                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
+                bool in_range = i < num_elements;
+                uint ii1;
+                if (nei0_is_pow2) {
+                    ii1 = i >> nei0shift;
+                } else {
+                    ii1 = i / p.nei0;
+                }
+                uint ii0 = i - ii1 * p.nei0;
+                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            }
+        }
+        uint i = j + gl_LocalInvocationIndex;
+        bool in_range = i < num_elements;
+        uint ii1;
+        if (nei0_is_pow2) {
+            ii1 = i >> nei0shift;
+        } else {
+            ii1 = i / p.nei0;
+        }
+        uint ii0 = i - ii1 * p.nei0;
+        uint id = ids[iter++];
+        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+
+        ballots_sh[gl_SubgroupID] = ballot;
+        barrier();
+
+        uint subgroup_base = 0;
+        uint total = 0;
+        for (uint k = 0; k < gl_NumSubgroups; ++k) {
+            if (k == gl_SubgroupID) {
+                subgroup_base = total;
+            }
+            total += subgroupBallotBitCount(ballots_sh[k]);
+        }
+        barrier();
+
+        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
+        if (in_range && id == expert_idx) {
+            row_ids[_ne1 + idx] = u16vec4(fastmod(ii0, p.ne11), ii1, ii0, 0);
+        }
+        _ne1 += total;
+        iter &= 15;
+    }
+    barrier();
+}
 #endif
 
 void main() {
@@ -157,45 +217,12 @@ void main() {
     const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
-    // Spread the search across all elements in the first subgroup
-    if (gl_SubgroupID == 0) {
-        _ne1 = 0;
-        uint num_elements = p.nei1 * p.nei0;
-
-        uint ids[16];
-        uint iter = 0;
-
-        for (uint j = 0; j < num_elements; j += gl_SubgroupSize) {
-            // prefetch up to 16 elements
-            if (iter == 0) {
-                [[unroll]] for (uint k = 0; k < 16; ++k) {
-                    uint i = j + gl_SubgroupInvocationID + k*gl_SubgroupSize;
-                    bool in_range = i < num_elements;
-                    uint ii1 = i / p.nei0;
-                    uint ii0 = i % p.nei0;
-                    ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
-                }
-            }
-            uint i = j + gl_SubgroupInvocationID;
-            bool in_range = i < num_elements;
-            uint ii1 = i / p.nei0;
-            uint ii0 = i % p.nei0;
-            uint id = ids[iter++];
-            uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
-            uint idx = subgroupBallotExclusiveBitCount(ballot);
-            if (in_range && id == expert_idx) {
-                row_ids[_ne1 + idx] = u16vec4(ii0 % p.ne11, ii1, ii0, 0);
-            }
-            _ne1 += subgroupBallotBitCount(ballot);
-            iter &= 15;
-        }
-        _ne1_sh = _ne1;
+    if (bitCount(p.nei0) == 1) {
+        load_row_ids(expert_idx, true);
+    } else {
+        load_row_ids(expert_idx, false);
     }
 
-    barrier();
-
-    _ne1 = _ne1_sh;
-
     // Workgroup has no work
     if (ic * BN >= _ne1) return;
 #endif
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp b/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
new file mode 100644
index 0000000000000..854a2ad818760
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp
@@ -0,0 +1,111 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_nonuniform_qualifier : enable
+#extension GL_EXT_control_flow_attributes : require
+#if ADD_RMS
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#endif
+
+#include "rte.comp"
+#include "types.comp"
+#include "utils.comp"
+
+layout (push_constant) uniform parameter2
+{
+    // shape for dst
+    uint ne20; uint ne21; uint ne22; uint ne23;
+
+    // strides for srcs+dst
+    uint nb[12][4];
+
+    uint rms_partials;
+} p;
+
+// Workaround for MoltenVK Bug, see https://github.com/ggml-org/llama.cpp/issues/15498
+// layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
+// layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
+layout (binding = 0) buffer A {A_TYPE data_a[];} a[];
+layout (binding = 0) buffer D {D_TYPE data_d[];} d[];
+
+layout (binding = 0, std430) buffer PartialBuf {float partial_sums[];} partials[];
+
+layout(constant_id = 0) const uint num_srcs = 2;
+
+uint src_idx(uint s, uint i00, uint i01, uint i02, uint i03) {
+    return i03*p.nb[s][3] + i02*p.nb[s][2] + i01*p.nb[s][1] + i00*p.nb[s][0];
+}
+
+uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
+    uint nb20 = p.nb[num_srcs][0];
+    uint nb21 = p.nb[num_srcs][1];
+    uint nb22 = p.nb[num_srcs][2];
+    uint nb23 = p.nb[num_srcs][3];
+    return i03*nb23 + i02*nb22 + i01*nb21 + i00*nb20;
+}
+
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+#if ADD_RMS
+// XXX TODO this could be sized based on number of subgroups, but that't not considered a constant
+shared FLOAT_TYPE sumsh[num_threads];
+#endif
+
+void main() {
+    uint idx = get_idx();
+    uint orig_idx = idx;
+
+    uint ne = p.ne20 * p.ne21 * p.ne22 * p.ne23;
+
+    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
+    const uint num_iter = 2;
+
+    FLOAT_TYPE sum_sq = 0;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03, p.ne20, p.ne21, p.ne22, p.ne23);
+
+        FLOAT_TYPE sum = FLOAT_TYPE(0);
+        [[unroll]] for (uint s = 0; s < num_srcs; ++s) {
+            sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
+        }
+        sum_sq += sum*sum;
+        d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
+
+        idx += num_threads;
+    }
+
+#if ADD_RMS
+    if (p.rms_partials != 0) {
+        // reduce the sum within each subgroup, then across subgroups
+        const uint NumSubgroups = num_threads / gl_SubgroupSize;
+        sum_sq = subgroupAdd(sum_sq);
+        if (gl_SubgroupInvocationID == 0) {
+            sumsh[gl_SubgroupID] = sum_sq;
+        }
+        barrier();
+        [[unroll]] for (uint s = NumSubgroups / 2; s > 0; s >>= 1) {
+            if (gl_SubgroupID < s && gl_SubgroupInvocationID == 0) {
+                sum_sq += sumsh[gl_SubgroupID + s];
+                sumsh[gl_SubgroupID] = sum_sq;
+            }
+            barrier();
+        }
+
+        if (gl_SubgroupID == 0 && gl_SubgroupInvocationID == 0) {
+            partials[num_srcs + 1].partial_sums[orig_idx / (num_iter * num_threads)] = sum_sq;
+        }
+    }
+#endif
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp b/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp
new file mode 100644
index 0000000000000..6426dedee57b0
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) buffer X {A_TYPE data_x[];};
+layout (binding = 1) readonly buffer G {A_TYPE data_grad[];};
+layout (binding = 2) readonly buffer P {float data_params[2];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float alpha = data_params[0];
+    const float keep = 1.f - alpha * data_params[1];
+
+    data_x[i] = data_x[i] * keep - alpha * data_grad[i];
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
index bdd7db2d6987a..41197e9301ad8 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
@@ -10,9 +10,9 @@ layout (constant_id = 1) const bool do_multiply = false;
 
 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
-shared FLOAT_TYPE sum[BLOCK_SIZE];
+shared FLOAT_TYPE sumsh[BLOCK_SIZE];
 
-void main() {
+void rms_norm(uint num_iters) {
     const uint ncols     = p.ne00;
     const uint nrows     = gl_NumWorkGroups.x;
     const uint nchannels = gl_NumWorkGroups.y;
@@ -30,38 +30,76 @@ void main() {
     uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
     uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
 
-    sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
-    [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
-        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[a_offset + col]);
-        sum[tid] += xi * xi;
+    [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
+        FLOAT_TYPE xi = FLOAT_TYPE(0);
+        if (col < ncols) {
+            xi = FLOAT_TYPE(data_a[a_offset + col]);
+        }
+        sum += xi * xi;
     }
 
+    sumsh[tid] = sum;
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) {
-            sum[tid] += sum[tid + s];
+            sum += sumsh[tid + s];
+            sumsh[tid] = sum;
         }
         barrier();
     }
+    sum = sumsh[0];
 
-    const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols);
+    const FLOAT_TYPE mean = sum / FLOAT_TYPE(ncols);
     const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
 
     if (do_multiply) {
         if (ncols > p.ne10) {
-            [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+            [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
+                if (col >= ncols) {
+                    continue;
+                }
                 data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
             }
         } else {
-            [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+            [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
+                if (col >= ncols) {
+                    continue;
+                }
                 data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
             }
         }
     } else {
-        [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+        [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
+            if (col >= ncols) {
+                continue;
+            }
             data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
         }
     }
 }
+
+void main() {
+    // instantiate the rms_norm function for several different
+    // dimensions, to allow loop unrolling
+    uint num_blocks = (p.ne00 + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    if (num_blocks > 32) {
+        rms_norm(num_blocks);
+    } else if (num_blocks > 16) {
+        rms_norm(32);
+    } else if (num_blocks > 8) {
+        rms_norm(16);
+    } else if (num_blocks > 4) {
+        rms_norm(8);
+    } else if (num_blocks == 4) {
+        rms_norm(4);
+    } else if (num_blocks == 3) {
+        rms_norm(3);
+    } else if (num_blocks == 2) {
+        rms_norm(2);
+    } else if (num_blocks == 1) {
+        rms_norm(1);
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp
new file mode 100644
index 0000000000000..ba4677c293392
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp
@@ -0,0 +1,65 @@
+#version 450
+
+#include "generic_binary_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+
+#define BLOCK_SIZE 128
+
+layout (constant_id = 1) const bool do_multiply = false;
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 3, std430) readonly buffer PartialsBuf {float partial_sums[];};
+
+shared FLOAT_TYPE sumsh[BLOCK_SIZE];
+
+void main() {
+    const uint ncols     = p.ne00;
+    const uint nrows     = gl_NumWorkGroups.x;
+    const uint nchannels = gl_NumWorkGroups.y;
+
+    const uint row       = 0;
+    const uint channel   = gl_WorkGroupID.y;
+    const uint samp      = gl_WorkGroupID.z;
+    // The work is split across multiple workgroups in the x dimension. Each invocation
+    // processes one element
+    const uint tid       = gl_GlobalInvocationID.x;
+
+    const uint stride_row       = p.nb01;
+    const uint stride_channel   = p.nb02;
+    const uint stride_sample    = p.nb03;
+
+    uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
+    uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
+    uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
+
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f); // partial sum for thread in warp
+
+    uint32_t num_partials = p.param3;
+    for (uint32_t i = gl_SubgroupInvocationID; i < num_partials; i += gl_SubgroupSize) {
+        sum += partial_sums[i];
+    }
+    sum = subgroupAdd(sum);
+
+    uint col = tid;
+    if (col >= ncols) {
+        return;
+    }
+
+    const FLOAT_TYPE mean = sum / FLOAT_TYPE(ncols);
+    const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
+
+    if (do_multiply) {
+        if (ncols > p.ne10) {
+            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
+        } else {
+            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+        }
+    } else {
+        data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp b/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp
new file mode 100644
index 0000000000000..4bc697b9b9c54
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp
@@ -0,0 +1,17 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sqrt(val));
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp b/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
index 961e5ffa1f56f..759204afaf9aa 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
@@ -1,9 +1,9 @@
 #version 450
 
-#include "generic_head.comp"
 #include "types.comp"
 
 #extension GL_EXT_control_flow_attributes : enable
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -11,16 +11,49 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 
+layout (push_constant) uniform parameter
+{
+    uint n_cols;
+    uint ne01, ne02;
+    uint nb01, nb02, nb03;
+    uint nb11, nb12, nb13;
+    float weight;
+    uint misalign_offsets;
+    uint ne0_12mp, ne0_12L;
+    uint ne0_1mp, ne0_1L;
+} p;
+
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
+
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
+    const float weight = p.weight;
+
+    const uint i03 = fastdiv(row, p.ne0_12mp, p.ne0_12L);
+    const uint i03_offset = i03 * p.ne01*p.ne02;
+    const uint i02 = fastdiv(row - i03_offset, p.ne0_1mp, p.ne0_1L);
+    const uint i01 = row - i03_offset - i02*p.ne01;
+
+    const uint src_idx = get_aoffset() + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+    const uint dst_idx = get_doffset() + i01 * p.nb11 + i02 * p.nb12 + i03 * p.nb13;
 
-    tmp[col] = FLOAT_TYPE(0.0f);
+    tmp[col] = FLOAT_TYPE(0.0);
 
-    for (uint i = col; i < p.KX; i += BLOCK_SIZE) {
-        tmp[col] += FLOAT_TYPE(data_a[row*p.KX + i]);
+    for (uint i = col; i < p.n_cols; i += BLOCK_SIZE) {
+        tmp[col] += FLOAT_TYPE(data_a[src_idx + i]);
     }
 
     barrier();
@@ -32,6 +65,6 @@ void main() {
     }
 
     if (col == 0) {
-        data_d[row] = D_TYPE(tmp[0]);
+        data_d[dst_idx] = D_TYPE(tmp[0] * weight);
     }
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp b/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp
new file mode 100644
index 0000000000000..dc4a1e6d96bab
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp
@@ -0,0 +1,25 @@
+#ifndef UTILS_COMP
+#define UTILS_COMP
+
+// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
+uint fastmod(uint a, uint b) {
+    if ((b & (b-1)) == 0) {
+        return a & (b-1);
+    }
+    return a % b;
+}
+
+uint fastdiv(uint a, uint b) {
+    return (a < b) ? 0 : (a / b);
+}
+
+void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03, uint ne00, uint ne01, uint ne02, uint ne03) {
+    i03 = fastdiv(idx, (ne02*ne01*ne00));
+    const uint i03_offset = i03 * ne02*ne01*ne00;
+    i02 = fastdiv((idx - i03_offset), (ne01*ne00));
+    const uint i02_offset = i02*ne01*ne00;
+    i01 = (idx - i03_offset - i02_offset) / ne00;
+    i00 = idx - i03_offset - i02_offset - i01*ne00;
+}
+
+#endif // UTILS_COMP
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 4cd94c51e3f21..50a27748317be 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -223,7 +223,8 @@ void string_to_spv_func(const std::string& _name, const std::string& in_fname, c
     std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2";
 
     // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
-    std::string opt_level = coopmat ? "" : "-O";
+    // disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
+    std::string opt_level = (coopmat || name.find("bf16") != std::string::npos) ? "" : "-O";
 
     #ifdef _WIN32
         std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
@@ -472,6 +473,9 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
         string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
 
+        string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+
         string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
@@ -499,6 +503,7 @@ void process_shaders() {
     string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
@@ -534,13 +539,15 @@ void process_shaders() {
         s += std::string(dst_f16 ? "_f16" : "_f32");
         return s;
     };
-    for (std::string op : {"add", "sub", "mul", "div"}) {
+    for (std::string op : {"add", "sub", "mul", "div", "add_rms", }) {
     for (auto src0_f16 : {false, true}) {
     for (auto src1_f16 : {false, true}) {
     for (auto dst_f16  : {false, true}) {
     for (auto rte      : {false, true}) {
+        auto source = op == "add_rms" ? std::string("add") : op;
         auto name = op + get_suffix(src0_f16, src1_f16, dst_f16) + (rte ? "_rte" : "");
-        string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
+        auto add_rms = op == "add_rms" ? "1" : "0";
+        string_to_spv(name.c_str(), source + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}, {"ADD_RMS" , add_rms}});
     }
     }
     }
@@ -566,6 +573,8 @@ void process_shaders() {
 
     string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
+    string_to_spv("sqrt_f32", "sqrt.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+
     string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
     string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
@@ -580,6 +589,8 @@ void process_shaders() {
 
     string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
 
+    string_to_spv("exp_f16",        "exp.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("exp_f32",        "exp.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_f16",       "gelu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("gelu_f32",       "gelu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_erf_f16",   "gelu_erf.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
@@ -657,6 +668,7 @@ void process_shaders() {
     string_to_spv("rwkv_wkv7_f32", "wkv7.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
+    string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
     string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
     string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
@@ -671,11 +683,16 @@ void process_shaders() {
 
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
+    string_to_spv("conv2d_dw_whcn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
+    string_to_spv("conv2d_dw_cwhn_f16_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
 
     string_to_spv("roll_f32", "roll.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}, {"ADD_RMS" , "0"}});
+    string_to_spv("multi_add_rms_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}, {"ADD_RMS" , "1"}});
+
     for (auto &c : compiles) {
         c.wait();
     }
@@ -732,7 +749,7 @@ void write_output_files() {
     }
 
     std::string suffixes[2] = {"_f32", "_f16"};
-    for (const char *op : {"add", "sub", "mul", "div"}) {
+    for (const char *op : {"add", "sub", "mul", "div", "add_rms"}) {
         fprintf(hdr, "extern unsigned char *%s_data[2][2][2][2];\n", op);
         fprintf(hdr, "extern uint64_t %s_len[2][2][2][2];\n", op);
         std::string data = "unsigned char *" + std::string(op) + "_data[2][2][2][2] = ";
@@ -784,6 +801,18 @@ void write_output_files() {
         fputs(data.c_str(), src);
         fputs(len.c_str(), src);
     }
+
+    for (const std::string& btype : {"f16", "f32"}) {
+    for (const auto& tname : type_names) {
+        fprintf(hdr, "extern unsigned char *arr_dmmv_%s_%s_f32_data[2];\n", tname.c_str(), btype.c_str());
+        fprintf(hdr, "extern uint64_t arr_dmmv_%s_%s_f32_len[2];\n", tname.c_str(), btype.c_str());
+        std::string data = "unsigned char *arr_dmmv_" + tname + "_" + btype + "_f32_data[2] = {mul_mat_vec_" + tname + "_" + btype + "_f32_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_data};\n";
+        std::string len =  "uint64_t arr_dmmv_"       + tname + "_" + btype + "_f32_len[2] =  {mul_mat_vec_" + tname + "_" + btype + "_f32_len,  mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_len};\n";
+        fputs(data.c_str(), src);
+        fputs(len.c_str(), src);
+    }
+    }
+
     fclose(hdr);
     fclose(src);
 }
diff --git a/ggml/src/ggml-webgpu/CMakeLists.txt b/ggml/src/ggml-webgpu/CMakeLists.txt
index 79ef68b85a477..78a985a4d167a 100644
--- a/ggml/src/ggml-webgpu/CMakeLists.txt
+++ b/ggml/src/ggml-webgpu/CMakeLists.txt
@@ -20,8 +20,8 @@ add_custom_command(
     COMMAND ${CMAKE_COMMAND} -E make_directory ${SHADER_OUTPUT_DIR}
     COMMAND ${CMAKE_COMMAND} -E env PYTHONIOENCODING=utf-8
         ${Python3_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
-            --input "${SHADER_DIR}"
-            --output "${SHADER_HEADER}"
+            --input_dir "${SHADER_DIR}"
+            --output_file "${SHADER_HEADER}"
     DEPENDS ${WGSL_SHADER_FILES} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
     VERBATIM
 )
diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
index ba1addc8d9f29..32f1e304e1e63 100644
--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -118,13 +118,11 @@ struct webgpu_context_struct {
 
     std::recursive_mutex mutex;
 
-    bool device_init = false;
-
     webgpu_buf_pool param_buf_pool;
     webgpu_buf_pool set_rows_error_buf_pool;
 
     wgpu::ComputePipeline memset_pipeline;
-    wgpu::ComputePipeline mul_mat_pipeline;
+    wgpu::ComputePipeline mul_mat_pipeline[30][2];
     wgpu::ComputePipeline set_rows_pipeline;
     wgpu::ComputePipeline cpy_pipeline;
 
@@ -238,7 +236,7 @@ static void ggml_backend_webgpu_wait_on_submission(webgpu_context & ctx) {
                                   wgpu::CallbackMode::AllowSpontaneous,
                                   [](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
                                       if (status != wgpu::QueueWorkDoneStatus::Success) {
-                                          GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", message.data);
+                                          GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str());
                                       }
                                   }),
                               UINT64_MAX);
@@ -278,7 +276,7 @@ static void ggml_backend_webgpu_submit_queue(webgpu_context & ctx) {
         wgpu::CallbackMode::AllowSpontaneous,
         [ctx, staged_param_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
             if (status != wgpu::QueueWorkDoneStatus::Success) {
-                GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", message.data);
+                GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str());
             }
             // Free the staged buffers
             ctx->param_buf_pool.free_bufs(staged_param_bufs);
@@ -294,7 +292,7 @@ static void ggml_backend_webgpu_submit_queue(webgpu_context & ctx) {
             wgpu::CallbackMode::AllowSpontaneous,
             [ctx, error_bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) {
                 if (status != wgpu::MapAsyncStatus::Success) {
-                    GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", message.data);
+                    GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", std::string(message).c_str());
                 } else {
                     const uint32_t * error_data = (const uint32_t *) error_bufs.host_buf.GetConstMappedRange();
                     if (*error_data) {
@@ -331,6 +329,7 @@ static void ggml_backend_webgpu_map_buffer(webgpu_context & ctx,
 // To use, add a bind group entry to the setup for the shader you are debugging, add the buffer and
 // debug statements in the shader, and then call this function after encoding the commands and submitting them.
 static void ggml_backend_webgpu_debug(webgpu_context & ctx) {
+    ggml_backend_webgpu_submit_queue(ctx);
     wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
     encoder.CopyBufferToBuffer(ctx->debug_dev_buf, 0, ctx->debug_host_buf, 0, ctx->debug_host_buf.GetSize());
     wgpu::CommandBuffer commands = encoder.Finish();
@@ -421,15 +420,6 @@ static void ggml_backend_webgpu_buffer_memset(webgpu_context & ctx,
     ggml_backend_webgpu_build_and_enqueue(ctx, ctx->memset_pipeline, params, entries, wg_x, true);
 }
 
-static size_t ggml_backend_webgpu_tensor_offset(const ggml_tensor * tensor) {
-    return webgpu_tensor_offset(tensor) + tensor->view_offs;
-}
-
-static wgpu::Buffer ggml_backend_webgpu_tensor_buf(const ggml_tensor * tensor) {
-    ggml_backend_webgpu_buffer_context * ctx = (ggml_backend_webgpu_buffer_context *) tensor->buffer->context;
-    return ctx->buffer;
-}
-
 /** End WebGPU Actions */
 
 /** GGML Backend Interface */
@@ -447,19 +437,36 @@ static void ggml_backend_webgpu_free(ggml_backend_t backend) {
     GGML_UNUSED(ctx);
 }
 
+static size_t ggml_webgpu_tensor_offset(const ggml_tensor * tensor) {
+    return webgpu_tensor_offset(tensor) + tensor->view_offs;
+}
+
+static wgpu::Buffer ggml_webgpu_tensor_buf(const ggml_tensor * tensor) {
+    ggml_backend_webgpu_buffer_context * ctx = (ggml_backend_webgpu_buffer_context *) tensor->buffer->context;
+    return ctx->buffer;
+}
+
+static size_t ggml_webgpu_tensor_misalignment(webgpu_context & ctx, ggml_tensor * t) {
+    size_t offset = ggml_webgpu_tensor_offset(t);
+    return offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+}
+
+static size_t ggml_webgpu_tensor_align_offset(webgpu_context & ctx, ggml_tensor * t) {
+    size_t offset = ggml_webgpu_tensor_offset(t);
+    return offset & ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+}
+
+static size_t ggml_webgpu_tensor_binding_size(webgpu_context & ctx, ggml_tensor * t) {
+    return (ggml_nbytes(t) + ggml_webgpu_tensor_misalignment(ctx, t) + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
+           ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1);
+}
+
 static void ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
-    size_t src_offset       = ggml_backend_webgpu_tensor_offset(src);
-    // assumes power of 2 offset alignment
-    size_t src_misalignment = src_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    // align to minimum offset alignment
-    src_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-    size_t dst_offset       = ggml_backend_webgpu_tensor_offset(dst);
-    size_t dst_misalignment = dst_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    dst_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-    uint32_t              ne     = (uint32_t) ggml_nelements(dst);
+    uint32_t ne = (uint32_t) ggml_nelements(dst);
+
     std::vector<uint32_t> params = { ne,
-                                     (uint32_t) (src_misalignment / ggml_type_size(src->type)),
-                                     (uint32_t) (dst_misalignment / ggml_type_size(dst->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
                                      // Convert byte-strides to element-strides
                                      (uint32_t) (src->nb[0] / ggml_type_size(src->type)),
                                      (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
@@ -477,15 +484,13 @@ static void ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, ggml_tensor
 
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0,
-         .buffer  = ggml_backend_webgpu_tensor_buf(src),
-         .offset  = src_offset,
-         .size    = (ggml_nbytes(src) + src_misalignment + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
-                  ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1) },
+         .buffer  = ggml_webgpu_tensor_buf(src),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src) },
         { .binding = 1,
-         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
-         .offset  = dst_offset,
-         .size    = (ggml_nbytes(dst) + dst_misalignment + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) &
-                  ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1) }
+         .buffer  = ggml_webgpu_tensor_buf(dst),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, dst) }
     };
 
     size_t   max_wg_size = ctx->limits.maxComputeWorkgroupSizeX;
@@ -504,21 +509,9 @@ static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_t
         error_bufs.host_buf.Unmap();
     }
 
-    size_t src_offset       = ggml_backend_webgpu_tensor_offset(src);
-    // assumes power of 2 offset alignment
-    size_t src_misalignment = src_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    // align to minimum offset alignment
-    src_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-    size_t idx_offset       = ggml_backend_webgpu_tensor_offset(idx);
-    size_t idx_misalignment = idx_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    idx_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-    size_t dst_offset       = ggml_backend_webgpu_tensor_offset(dst);
-    size_t dst_misalignment = dst_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
-    dst_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
-
-    std::vector<uint32_t> params = { (uint32_t) (src_misalignment / ggml_type_size(src->type)),
-                                     (uint32_t) (idx_misalignment / ggml_type_size(idx->type)),
-                                     (uint32_t) (dst_misalignment / ggml_type_size(dst->type)),
+    std::vector<uint32_t> params = { (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src) / ggml_type_size(src->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, idx) / ggml_type_size(idx->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
                                      // Convert byte-strides to element-strides
                                      (uint32_t) (src->nb[1] / ggml_type_size(src->type)),
                                      (uint32_t) (src->nb[2] / ggml_type_size(src->type)),
@@ -540,18 +533,18 @@ static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_t
 
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0,
-         .buffer  = ggml_backend_webgpu_tensor_buf(src),
-         .offset  = ggml_backend_webgpu_tensor_offset(src),
-         .size    = ggml_nbytes(src)                                                                       },
+         .buffer  = ggml_webgpu_tensor_buf(src),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src) },
         { .binding = 1,
-         .buffer  = ggml_backend_webgpu_tensor_buf(idx),
-         .offset  = ggml_backend_webgpu_tensor_offset(idx),
-         .size    = ggml_nbytes(idx)                                                                       },
+         .buffer  = ggml_webgpu_tensor_buf(idx),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, idx),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, idx) },
         { .binding = 2,
-         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
-         .offset  = ggml_backend_webgpu_tensor_offset(dst),
-         .size    = ggml_nbytes(dst)                                                                       },
-        { .binding = 3, .buffer = error_bufs.dev_buf,    .offset = 0, .size = error_bufs.dev_buf.GetSize() }
+         .buffer  = ggml_webgpu_tensor_buf(dst),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, dst) },
+        { .binding = 3, .buffer = error_bufs.dev_buf, .offset = 0, .size = error_bufs.dev_buf.GetSize() }
     };
 
     size_t   max_wg_size = ctx->limits.maxComputeWorkgroupSizeX;
@@ -565,15 +558,18 @@ static void ggml_webgpu_set_rows(webgpu_context & ctx, ggml_tensor * src, ggml_t
 
 static void ggml_webgpu_mul_mat(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst) {
     std::vector<uint32_t> params = {
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
         (uint32_t) dst->ne[1],                                  // number of rows in result (M)
         (uint32_t) dst->ne[0],                                  // number of columns in result (N)
         (uint32_t) src0->ne[0],                                 // number of columns in src0/src1 (K)
-        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 1
-        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 1
-        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 2
-        (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 2
-        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),  // stride (elements) of src0 in dimension 3
-        (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),  // stride (elements) of src1 in dimension 3
+        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),  // stride (elements/blocks) of src0 in dimension 1
+        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),  // stride (elements/blocks) of src1 in dimension 1
+        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),  // stride (elements/blocks) of src0 in dimension 2
+        (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),  // stride (elements/blocks) of src1 in dimension 2
+        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),  // stride (elements/blocks) of src0 in dimension 3
+        (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),  // stride (elements/blocks) of src1 in dimension 3
         (uint32_t) src0->ne[2],                                 // batch size in dimension 2
         (uint32_t) src0->ne[3],                                 // batch size in dimension 3
         (uint32_t) (src1->ne[2] / src0->ne[2]),                 // broadcast in dimension 2
@@ -582,22 +578,22 @@ static void ggml_webgpu_mul_mat(webgpu_context & ctx, ggml_tensor * src0, ggml_t
 
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0,
-         .buffer  = ggml_backend_webgpu_tensor_buf(src0),
-         .offset  = ggml_backend_webgpu_tensor_offset(src0),
-         .size    = ggml_nbytes(src0) },
+         .buffer  = ggml_webgpu_tensor_buf(src0),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src0),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src0) },
         { .binding = 1,
-         .buffer  = ggml_backend_webgpu_tensor_buf(src1),
-         .offset  = ggml_backend_webgpu_tensor_offset(src1),
-         .size    = ggml_nbytes(src1) },
+         .buffer  = ggml_webgpu_tensor_buf(src1),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src1),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src1) },
         { .binding = 2,
-         .buffer  = ggml_backend_webgpu_tensor_buf(dst),
-         .offset  = ggml_backend_webgpu_tensor_offset(dst),
-         .size    = ggml_nbytes(dst)  }
+         .buffer  = ggml_webgpu_tensor_buf(dst),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, dst)  },
     };
 
     uint32_t wg_x =
         (dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3] + WEBGPU_MUL_MAT_WG_SIZE - 1) / WEBGPU_MUL_MAT_WG_SIZE;
-    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->mul_mat_pipeline, params, entries, wg_x);
+    ggml_backend_webgpu_build_and_enqueue(ctx, ctx->mul_mat_pipeline[src0->type][src1->type], params, entries, wg_x);
 }
 
 // Returns true if node has enqueued work into the queue, false otherwise
@@ -827,7 +823,7 @@ static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_b
     wgpu::Buffer buf;
     ggml_webgpu_create_buffer(ctx->webgpu_ctx->device,
                               buf,
-                              size,
+                              (size + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) & ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1),
                               wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst,
                               "allocated_buffer");
 
@@ -907,7 +903,94 @@ static void ggml_webgpu_init_memset_pipeline(webgpu_context & webgpu_ctx) {
 }
 
 static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline, wgsl_mul_mat, "mul_mat");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F32][GGML_TYPE_F32],
+                                wgsl_mul_mat_f32_f32,
+                                "mul_mat_f32_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F16][GGML_TYPE_F16],
+                                wgsl_mul_mat_f16_f16,
+                                "mul_mat_f16_f16");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F16][GGML_TYPE_F32],
+                                wgsl_mul_mat_f16_f32,
+                                "mul_mat_f16_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_0][GGML_TYPE_F32],
+                                wgsl_mul_mat_q4_0_f32,
+                                "mul_mat_q4_0_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_1][GGML_TYPE_F32],
+                                wgsl_mul_mat_q4_1_f32,
+                                "mul_mat_q4_1_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_0][GGML_TYPE_F32],
+                                wgsl_mul_mat_q5_0_f32,
+                                "mul_mat_q5_0_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_1][GGML_TYPE_F32],
+                                wgsl_mul_mat_q5_1_f32,
+                                "mul_mat_q5_1_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q8_0][GGML_TYPE_F32],
+                                wgsl_mul_mat_q8_0_f32,
+                                "mul_mat_q8_0_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q2_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q2_k_f32,
+                                "mul_mat_q2_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q3_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q3_k_f32,
+                                "mul_mat_q3_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q4_k_f32,
+                                "mul_mat_q4_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q5_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q5_k_f32,
+                                "mul_mat_q5_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q6_K][GGML_TYPE_F32],
+                                wgsl_mul_mat_q6_k_f32,
+                                "mul_mat_q6_k_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_XXS][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq2_xxs_f32,
+                                "mul_mat_iq2_xxs_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_XS][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq2_xs_f32,
+                                "mul_mat_iq2_xs_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ2_S][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq2_s_f32,
+                                "mul_mat_iq2_s_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ3_XXS][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq3_xxs_f32,
+                                "mul_mat_iq3_xxs_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ3_S][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq3_s_f32,
+                                "mul_mat_iq3_s_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ1_S][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq1_s_f32,
+                                "mul_mat_iq1_s_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ1_M][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq1_m_f32,
+                                "mul_mat_iq1_m_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ4_NL][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq4_nl_f32,
+                                "mul_mat_iq4_nl_f32");
+    ggml_webgpu_create_pipeline(webgpu_ctx->device,
+                                webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ4_XS][GGML_TYPE_F32],
+                                wgsl_mul_mat_iq4_xs_f32,
+                                "mul_mat_iq4_xs_f32");
 }
 
 static void ggml_webgpu_init_set_rows_pipeline(webgpu_context & webgpu_ctx) {
@@ -933,79 +1016,6 @@ static ggml_backend_t ggml_backend_webgpu_device_init(ggml_backend_dev_t dev, co
     ggml_backend_webgpu_device_context * dev_ctx    = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
     webgpu_context                       webgpu_ctx = dev_ctx->webgpu_ctx;
 
-    // Multiple threads may try to initialize the device
-    std::lock_guard<std::recursive_mutex> lock(webgpu_ctx->mutex);
-    if (!webgpu_ctx->device_init) {
-        // Initialize device
-        std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16,
-                                                             wgpu::FeatureName::ImplicitDeviceSynchronization };
-        wgpu::DeviceDescriptor         dev_desc;
-        dev_desc.requiredLimits       = &webgpu_ctx->limits;
-        dev_desc.requiredFeatures     = required_features.data();
-        dev_desc.requiredFeatureCount = required_features.size();
-        dev_desc.SetDeviceLostCallback(
-            wgpu::CallbackMode::AllowSpontaneous,
-            [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
-                GGML_UNUSED(device);
-                GGML_LOG_ERROR(
-                    "ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason), message.data);
-            });
-        dev_desc.SetUncapturedErrorCallback(
-            [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
-                GGML_UNUSED(device);
-                GGML_LOG_ERROR(
-                    "ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason), message.data);
-            });
-        webgpu_ctx->instance.WaitAny(
-            webgpu_ctx->adapter.RequestDevice(
-                &dev_desc,
-                wgpu::CallbackMode::AllowSpontaneous,
-                [webgpu_ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
-                    if (status != wgpu::RequestDeviceStatus::Success) {
-                        GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", message.data);
-                        return;
-                    }
-                    webgpu_ctx->device = std::move(device);
-                }),
-            UINT64_MAX);
-        GGML_ASSERT(webgpu_ctx->device != nullptr);
-
-        // Initialize (compute) queue
-        webgpu_ctx->queue = webgpu_ctx->device.GetQueue();
-
-        // Create buffer pool for shader parameters
-        webgpu_ctx->param_buf_pool.init(webgpu_ctx->device,
-                                        WEBGPU_NUM_PARAM_BUFS,
-                                        WEBGPU_PARAMS_BUF_SIZE_BYTES,
-                                        wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
-                                        wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
-        webgpu_ctx->set_rows_error_buf_pool.init(webgpu_ctx->device,
-                                                 WEBGPU_NUM_SET_ROWS_ERROR_BUFS,
-                                                 WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES,
-                                                 wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage,
-                                                 wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead);
-
-        ggml_webgpu_init_memset_pipeline(webgpu_ctx);
-        ggml_webgpu_init_mul_mat_pipeline(webgpu_ctx);
-        ggml_webgpu_init_set_rows_pipeline(webgpu_ctx);
-        ggml_webgpu_init_cpy_pipeline(webgpu_ctx);
-
-#ifdef GGML_WEBGPU_DEBUG
-        // Initialize debug buffers
-        ggml_webgpu_create_buffer(webgpu_ctx->device,
-                                  webgpu_ctx->debug_host_buf,
-                                  WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
-                                  wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead,
-                                  "debug_host_buf");
-        ggml_webgpu_create_buffer(webgpu_ctx->device,
-                                  webgpu_ctx->debug_dev_buf,
-                                  WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
-                                  wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc,
-                                  "debug_dev_buf");
-#endif
-        webgpu_ctx->device_init = true;
-    }
-
     static ggml_backend_webgpu_context backend_ctx;
     backend_ctx.name       = GGML_WEBGPU_NAME + std::string(": ") + dev_ctx->device_name;
     backend_ctx.webgpu_ctx = webgpu_ctx;
@@ -1053,10 +1063,45 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
             return true;
-        case GGML_OP_CPY | GGML_OP_SET_ROWS:
+        case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
             return op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_MUL_MAT:
-            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+            {
+                switch (op->src[1]->type) {
+                    case GGML_TYPE_F16:
+                        return op->src[0]->type == GGML_TYPE_F16;
+                    case GGML_TYPE_F32:
+                        switch (op->src[0]->type) {
+                            case GGML_TYPE_F32:
+                            case GGML_TYPE_F16:
+                            case GGML_TYPE_Q4_0:
+                            case GGML_TYPE_Q4_1:
+                            case GGML_TYPE_Q5_0:
+                            case GGML_TYPE_Q5_1:
+                            case GGML_TYPE_Q8_0:
+                            case GGML_TYPE_Q2_K:
+                            case GGML_TYPE_Q3_K:
+                            case GGML_TYPE_Q4_K:
+                            case GGML_TYPE_Q5_K:
+                            case GGML_TYPE_Q6_K:
+                            case GGML_TYPE_IQ2_XXS:
+                            case GGML_TYPE_IQ2_XS:
+                            case GGML_TYPE_IQ2_S:
+                            case GGML_TYPE_IQ3_XXS:
+                            case GGML_TYPE_IQ3_S:
+                            case GGML_TYPE_IQ1_S:
+                            case GGML_TYPE_IQ1_M:
+                            case GGML_TYPE_IQ4_NL:
+                            case GGML_TYPE_IQ4_XS:
+                                return true;
+                            default:
+                                return false;
+                        }
+                    default:
+                        return false;
+                }
+            }
         default:
             return false;
     }
@@ -1123,20 +1168,87 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
     wgpu::AdapterInfo info{};
     ctx->adapter.GetInfo(&info);
 
+    // Initialize device
+    std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16,
+                                                         wgpu::FeatureName::ImplicitDeviceSynchronization };
+    wgpu::DeviceDescriptor         dev_desc;
+    dev_desc.requiredLimits       = &ctx->limits;
+    dev_desc.requiredFeatures     = required_features.data();
+    dev_desc.requiredFeatureCount = required_features.size();
+    dev_desc.SetDeviceLostCallback(
+        wgpu::CallbackMode::AllowSpontaneous,
+        [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
+            GGML_UNUSED(device);
+            GGML_LOG_ERROR(
+                "ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason), std::string(message).c_str());
+        });
+    dev_desc.SetUncapturedErrorCallback(
+        [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
+            GGML_UNUSED(device);
+            GGML_LOG_ERROR(
+                "ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason), std::string(message).c_str());
+        });
+    ctx->instance.WaitAny(ctx->adapter.RequestDevice(
+                              &dev_desc,
+                              wgpu::CallbackMode::AllowSpontaneous,
+                              [ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
+                                  if (status != wgpu::RequestDeviceStatus::Success) {
+                                      GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", std::string(message).c_str());
+                                      return;
+                                  }
+                                  ctx->device = std::move(device);
+                              }),
+                          UINT64_MAX);
+    GGML_ASSERT(ctx->device != nullptr);
+
+    // Initialize (compute) queue
+    ctx->queue = ctx->device.GetQueue();
+
+    // Create buffer pool for shader parameters
+    ctx->param_buf_pool.init(ctx->device,
+                             WEBGPU_NUM_PARAM_BUFS,
+                             WEBGPU_PARAMS_BUF_SIZE_BYTES,
+                             wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
+                             wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
+    ctx->set_rows_error_buf_pool.init(ctx->device,
+                                      WEBGPU_NUM_SET_ROWS_ERROR_BUFS,
+                                      WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES,
+                                      wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage,
+                                      wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead);
+
+    ggml_webgpu_init_memset_pipeline(ctx);
+    ggml_webgpu_init_mul_mat_pipeline(ctx);
+    ggml_webgpu_init_set_rows_pipeline(ctx);
+    ggml_webgpu_init_cpy_pipeline(ctx);
+
+#ifdef GGML_WEBGPU_DEBUG
+    // Initialize debug buffers
+    ggml_webgpu_create_buffer(ctx->device,
+                              ctx->debug_host_buf,
+                              WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
+                              wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead,
+                              "debug_host_buf");
+    ggml_webgpu_create_buffer(ctx->device,
+                              ctx->debug_dev_buf,
+                              WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
+                              wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc,
+                              "debug_dev_buf");
+#endif
+
     static ggml_backend_webgpu_device_context device_ctx;
     device_ctx.webgpu_ctx  = ctx;
     device_ctx.device_name = GGML_WEBGPU_NAME;
-    device_ctx.device_desc = std::string(info.description.data);
+    device_ctx.device_desc = info.description;
 
     GGML_LOG_INFO(
         "ggml_webgpu: adapter_info: vendor_id: %u | vendor: %s | architecture: %s | device_id: %u | name: %s | "
         "device_desc: %s\n",
         info.vendorID,
-        info.vendor.data,
-        info.architecture.data,
+        std::string(info.vendor).c_str(),
+        std::string(info.architecture).c_str(),
         info.deviceID,
-        info.device.data,
-        info.description.data);
+        std::string(info.device).c_str(),
+        std::string(info.description).c_str());
 
     // See GGML Backend Device Interface section
     static ggml_backend_device device = {
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py b/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py
index 962dcd6b170ed..cc8def7f13ea4 100755
--- a/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py
@@ -1,35 +1,85 @@
 import os
+import re
+import ast
 import argparse
 
 
-def escape_triple_quotes(wgsl):
-    # Simple defense in case of embedded """
-    return wgsl.replace('"""', '\\"""')
+def extract_block(text, name):
+    pattern = rf'#define\({name}\)\s*(.*?)#end\({name}\)'
+    match = re.search(pattern, text, re.DOTALL)
+    if not match:
+        raise ValueError(f"Missing block: {name}")
+    return match.group(1).strip()
 
 
-def to_cpp_string_literal(varname, content):
-    return f'const char* wgsl_{varname} = R"({content})";\n'
+def parse_decls(decls_text):
+    decls = {}
+    for name, code in re.findall(r'#decl\((.*?)\)\s*(.*?)#enddecl\(\1\)', decls_text, re.DOTALL):
+        decls[name.strip()] = code.strip()
+    return decls
+
+
+def replace_placeholders(shader_text, replacements):
+    for key, val in replacements.items():
+        # Match {{KEY}} literally, where KEY is escaped
+        pattern = r'{{\s*' + re.escape(key) + r'\s*}}'
+        shader_text = re.sub(pattern, str(val), shader_text)
+    return shader_text
+
+
+def write_shader(shader_name, shader_code, output_dir, outfile):
+    if output_dir:
+        wgsl_filename = os.path.join(output_dir, f"{shader_name}.wgsl")
+        with open(wgsl_filename, "w", encoding="utf-8") as f_out:
+            f_out.write(shader_code)
+    outfile.write(f'const char* wgsl_{shader_name} = R"({shader_code})";\n\n')
+
+
+def generate_variants(shader_path, output_dir, outfile):
+    shader_base_name = shader_path.split("/")[-1].split(".")[0]
+
+    with open(shader_path, "r", encoding="utf-8") as f:
+        text = f.read()
+
+    try:
+        variants = ast.literal_eval(extract_block(text, "VARIANTS"))
+    except ValueError:
+        write_shader(shader_base_name, text, output_dir, outfile)
+    else:
+        decls_map = parse_decls(extract_block(text, "DECLS"))
+        shader_template = extract_block(text, "SHADER")
+
+        for variant in variants:
+            decls = variant["DECLS"]
+            decls_code = ""
+            for key in decls:
+                if key not in decls_map:
+                    raise ValueError(f"DECLS key '{key}' not found.")
+                decls_code += decls_map[key] + "\n\n"
+
+            shader_variant = replace_placeholders(shader_template, variant["REPLS"])
+            final_shader = re.sub(r'\bDECLS\b', decls_code, shader_variant)
+
+            output_name = f"{shader_base_name}_" + "_".join([variant["REPLS"]["SRC0_TYPE"], variant["REPLS"]["SRC1_TYPE"]])
+            write_shader(output_name, final_shader, output_dir, outfile)
 
 
 def main():
     parser = argparse.ArgumentParser()
-    parser.add_argument('--input', required=True)
-    parser.add_argument('--output', required=True)
+    parser.add_argument("--input_dir", required=True)
+    parser.add_argument("--output_file", required=True)
+    parser.add_argument("--output_dir")
     args = parser.parse_args()
 
-    with open(args.output, 'w', encoding='utf-8') as out:
-        out.write("// Auto-generated shader embedding \n\n")
-        for fname in sorted(os.listdir(args.input)):
-            if not fname.endswith('.wgsl'):
-                continue
-            shader_path = os.path.join(args.input, fname)
-            varname = os.path.splitext(fname)[0]
-            with open(shader_path, 'r', encoding='utf-8') as f:
-                content = f.read()
-            content = escape_triple_quotes(content)
-            out.write(to_cpp_string_literal(varname, content))
-            out.write('\n')
-
-
-if __name__ == '__main__':
+    if args.output_dir:
+        os.makedirs(args.output_dir, exist_ok=True)
+
+    with open(args.output_file, "w", encoding="utf-8") as out:
+        out.write("// Auto-generated shader embedding\n\n")
+        for fname in sorted(os.listdir(args.input_dir)):
+            if fname.endswith(".wgsl"):
+                generate_variants(os.path.join(args.input_dir, fname), args.output_dir, out)
+
+
+if __name__ == "__main__":
     main()
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl
index cb7c8c3e09e91..194d2d6f58c77 100644
--- a/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl
@@ -19,20 +19,20 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let start = params.offset;
     let end = params.offset + params.size;
 
-    for (var j: u32 = 0u; j < bytes_per_thread; j = j + 1u) {
+    for (var j: u32 = 0u; j < bytes_per_thread; j += 4) {
         let byte_index = start + i + j;
-        if (byte_index + 4u <= end) {
-            output_buffer[(byte_index >> 2u)] = params.value;
+        if (byte_index + 4 <= end) {
+            output_buffer[byte_index >> 2] = params.value;
         } else {
             // Handle tail (unaligned)
-            for (var k: u32 = 0u; k < 4u; k = k + 1u) {
+            for (var k: u32 = 0; k < 4; k++) {
                 let idx = byte_index + k;
                 if (idx < end) {
-                    let word_idx = idx >> 2u;
-                    let byte_offset = (idx & 3u) * 8u;
-                    let mask = ~(0xffu << byte_offset);
+                    let word_idx = idx >> 2;
+                    let bit_offset = (idx & 3) * 8u;
+                    let mask = ~(0xffu << bit_offset);
                     let existing = output_buffer[word_idx];
-                    output_buffer[word_idx] = (existing & mask) | ((params.value & 0xffu) << byte_offset);
+                    output_buffer[word_idx] = (existing & mask) | (params.value & (0xffu << bit_offset));
                 }
             }
         }
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl
new file mode 100644
index 0000000000000..79465c298d726
--- /dev/null
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl
@@ -0,0 +1,1794 @@
+#define(VARIANTS)
+
+[
+  {
+    "REPLS": {
+      "SRC0_TYPE" : "f32",
+      "SRC1_TYPE" : "f32",
+      "BLOCK_SIZE" : 1
+    },
+    "DECLS" : ["FLOAT"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f16",
+      "BLOCK_SIZE" : 1
+    },
+    "DECLS" : ["FLOAT"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f32",
+      "BLOCK_SIZE" : 1
+    },
+    "DECLS" : ["FLOAT"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q4_0",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 32
+    },
+    "DECLS": ["BYTE_HELPERS", "Q4_0"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q4_1",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 32
+    },
+    "DECLS": ["BYTE_HELPERS", "Q4_1"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q5_0",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 32
+    },
+    "DECLS": ["BYTE_HELPERS", "Q5_0"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q5_1",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 32
+    },
+    "DECLS": ["BYTE_HELPERS", "Q5_1"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q8_0",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 32
+    },
+    "DECLS": ["BYTE_HELPERS", "Q8_0"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q2_k",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "Q2_K"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q3_k",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "Q3_K"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q4_k",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q4_K"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q5_k",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q5_K"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "q6_k",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "Q6_K"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq2_xxs",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XXS"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq2_xs",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XS"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq2_s",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_S"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq3_xxs",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_XSS"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq3_s",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_S"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq1_s",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ1_TABLE","IQ1_S"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq1_m",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ1_TABLE","IQ1_M"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq4_nl",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 32,
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ4_TABLE", "IQ4_NL"]
+  },
+  {
+    "REPLS": {
+      "SRC0_TYPE": "iq4_xs",
+      "SRC1_TYPE": "f32",
+      "BLOCK_SIZE": 256,
+    },
+    "DECLS": ["BYTE_HELPERS", "IQ4_TABLE", "IQ4_XS"]
+  }
+]
+
+#end(VARIANTS)
+
+#define(DECLS)
+
+#decl(BYTE_HELPERS)
+
+fn get_byte(value: u32, index: u32) -> u32 {
+    return (value >> (index * 8)) & 0xFF;
+}
+
+fn get_byte_i32(value: u32, index: u32) -> i32 {
+    return bitcast<i32>(((value >> (index * 8)) & 0xFF) << 24) >> 24;
+}
+
+#enddecl(BYTE_HELPERS)
+
+#decl(FLOAT)
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    return f32(src0[src0_idx_base + offset]) * f32(src1[src1_idx_base + offset]);
+}
+#enddecl(FLOAT)
+
+#decl(Q4_0)
+struct q4_0 {
+    d: f16,
+    qs: array<f16, 8>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block_q4_0 = src0[src0_idx_base + offset];
+    let d = f32(block_q4_0.d);
+    var sum: f32 = 0.0;
+    for (var j: u32 = 0; j < 4; j++) {
+        let q_packed = bitcast<u32>(vec2(block_q4_0.qs[2 * j], block_q4_0.qs[2 * j + 1]));
+        for (var k: u32 = 0; k < 4; k++) {
+            let q_byte = get_byte(q_packed, k);
+            let q_hi = (f32((q_byte >> 4) & 0xF) - 8.0f) * d;
+            let q_lo = (f32(q_byte & 0xF) - 8.0f) * d;
+            let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
+            sum += q_lo * f32(src1[src1_offset]);
+            sum += q_hi * f32(src1[src1_offset + 16]);
+        }
+    }
+    return sum;
+}
+#enddecl(Q4_0)
+
+#decl(Q4_1)
+struct q4_1 {
+    d: f16,
+    m: f16,
+    qs: array<u32, 4>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block_q4_1 = src0[src0_idx_base + offset];
+    let d = f32(block_q4_1.d);
+    let m = f32(block_q4_1.m);
+    var sum: f32 = 0.0;
+    for (var j: u32 = 0; j < 4; j++) {
+        let q_packed = block_q4_1.qs[j];
+        for (var k: u32 = 0; k < 4; k++) {
+            let q_byte = get_byte(q_packed, k);
+            let q_hi = f32((q_byte >> 4) & 0xF) * d + m;
+            let q_lo = f32(q_byte & 0xF) * d + m;
+            let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
+            sum += q_lo * f32(src1[src1_offset]);
+            sum += q_hi * f32(src1[src1_offset + 16]);
+        }
+    }
+    return sum;
+}
+#enddecl(Q4_1)
+
+#decl(Q5_0)
+struct q5_0 {
+    d: f16,
+    qh: array<f16, 2>,
+    qs: array<f16, 8>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block_q5_0 = src0[src0_idx_base + offset];
+    let d = f32(block_q5_0.d);
+    var sum: f32 = 0.0;
+    let qh_packed = bitcast<u32>(vec2(block_q5_0.qh[0], block_q5_0.qh[1]));
+    for (var j: u32 = 0; j < 4; j++) {
+        let q_packed = bitcast<u32>(vec2(block_q5_0.qs[2 * j], block_q5_0.qs[2 * j + 1]));
+        for (var k: u32 = 0; k < 4; k++) {
+            let q_byte = get_byte(q_packed, k);
+            let qh_hi = (qh_packed >> (j * 4 + k + 12)) & 0x10;
+            let q_hi = (f32(((q_byte >> 4) & 0xF) | qh_hi) - 16.0) * d;
+            let qh_lo = ((qh_packed >> (j * 4 + k)) << 4) & 0x10;
+            let q_lo = (f32((q_byte & 0xF) | qh_lo) - 16.0) * d;
+            let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
+            sum += q_lo * f32(src1[src1_offset]);
+            sum += q_hi * f32(src1[src1_offset + 16]);
+        }
+    }
+    return sum;
+}
+#enddecl(Q5_0)
+
+#decl(Q5_1)
+struct q5_1 {
+    d: f16,
+    m: f16,
+    qh: u32,
+    qs: array<u32, 4>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block_q5_1 = src0[src0_idx_base + offset];
+    let d = f32(block_q5_1.d);
+    let m = f32(block_q5_1.m);
+    var sum: f32 = 0.0;
+    for (var j: u32 = 0; j < 4; j++) {
+        let q_packed = block_q5_1.qs[j];
+        for (var k: u32 = 0; k < 4; k++) {
+            let q_byte = get_byte(q_packed, k);
+            let qh_hi = (block_q5_1.qh >> (j * 4 + k + 12)) & 0x10;
+            let q_hi = f32(((q_byte >> 4) & 0xF) | qh_hi) * d + m;
+            let qh_lo = ((block_q5_1.qh >> (j * 4 + k)) << 4) & 0x10;
+            let q_lo = f32((q_byte & 0xF) | qh_lo) * d + m;
+            let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
+            sum += q_lo * f32(src1[src1_offset]);
+            sum += q_hi * f32(src1[src1_offset + 16]);
+        }
+    }
+    return sum;
+}
+#enddecl(Q5_1)
+
+#decl(Q8_0)
+struct q8_0 {
+    d: f16,
+    qs: array<f16, 16>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block_q8_0 = src0[src0_idx_base + offset];
+    let d = f32(block_q8_0.d);
+    var sum: f32 = 0.0;
+    for (var j: u32 = 0; j < 8; j++) {
+        let q_packed = bitcast<u32>(vec2(block_q8_0.qs[2 * j], block_q8_0.qs[2 * j + 1]));
+        for (var k: u32 = 0; k < 4; k++) {
+            let q_byte = get_byte_i32(q_packed, k);
+            let q_val = f32(q_byte) * d;
+            let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
+            sum += q_val * f32(src1[src1_offset]);
+        }
+    }
+    return sum;
+}
+#enddecl(Q8_0)
+
+#decl(Q8_1)
+struct q8_1 {
+    d: f16,
+    m: f16,
+    qs: array<u32, 8>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block_q8_1 = src0[src0_idx_base + offset];
+    let d = f32(block_q8_1.d);
+    let m = f32(block_q8_1.m);
+    var sum: f32 = 0.0;
+    for (var j: u32 = 0; j < 8; j++) {
+        let q_packed = block_q8_1.qs[j];
+        for (var k: u32 = 0; k < 4; k++) {
+            let q_byte = get_byte_i32(q_packed, k);
+            let q_val = f32(q_byte) * d + m;
+            let src1_offset = src1_idx_base + offset * 32 + j * 4 + k;
+            sum += q_val * f32(src1[src1_offset]);
+        }
+    }
+    return sum;
+}
+#enddecl(Q8_1)
+
+#decl(Q2_K)
+// 16 blocks of 16 elements each
+struct q2_k {
+    scales: array<u32, 4>,
+    qs: array<u32, 16>,
+    d: f16,
+    dmin: f16
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    let m = f32(block.dmin);
+    var sum = 0.0;
+    var src1_i = src1_idx_base + offset * 256;
+    var is: u32 = 0;
+    // 2 halves of the block (128 elements each)
+    for (var q_b_idx: u32 = 0; q_b_idx < 64; q_b_idx += 32) {
+        // 4 groups (each group has 2 blocks of 16 elements)
+        for (var shift: u32 = 0; shift < 8; shift += 2) {
+            // 2 blocks
+            for (var k: u32 = 0; k < 32; k += 16) {
+                let sc = get_byte(block.scales[is / 4], is % 4);
+                is++;
+                let dl = d * f32(sc & 0xF);
+                let ml = m * f32(sc >> 4);
+                for (var l: u32 = 0u; l < 16; l++) {
+                    let q_idx = q_b_idx + k + l;
+                    let q_byte = get_byte(block.qs[q_idx / 4], q_idx % 4);
+                    let qs_val = (q_byte >> shift) & 3;
+                    sum += (f32(qs_val) * dl - ml) * src1[src1_i];
+                    src1_i++;
+                }
+            }
+        }
+    }
+    return sum;
+}
+
+#enddecl(Q2_K)
+
+#decl(Q3_K)
+// 16 blocks of 16 elements each
+struct q3_k {
+    hmask: array<f16, 16>,
+    qs: array<f16, 32>,
+    scales: array<f16, 6>, // 6-bit quantized values
+    d: f16
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+
+    // extract 6-bit scales, which consist of 4-bits from first 8 bytes of scale,
+    // and 2-bits from the last 4 bytes
+    let kmask1: u32 = 0x03030303;
+    let kmask2: u32 = 0x0f0f0f0f;
+    var scale_vals: array<u32, 4>;
+    for (var i: u32 = 0; i < 4; i++) {
+        scale_vals[i] = bitcast<u32>(vec2(block.scales[2 * i], block.scales[2 * i + 1]));
+    }
+    var tmp: u32 = scale_vals[2];
+    scale_vals[2] = ((scale_vals[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+    scale_vals[3] = ((scale_vals[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+    scale_vals[0] = (scale_vals[0] & kmask2) | ((tmp & kmask1) << 4);
+    scale_vals[1] = (scale_vals[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+
+    // convert arrays of f16 -> u32
+    var hmask_vals: array<u32, 8>;
+    for (var i: u32 = 0; i < 8; i++) {
+        hmask_vals[i] = bitcast<u32>(vec2(block.hmask[2 * i], block.hmask[2 * i + 1]));
+    }
+    var qs_vals: array<u32, 16>;
+    for (var i: u32 = 0; i < 16; i++) {
+        qs_vals[i] = bitcast<u32>(vec2(block.qs[2 * i], block.qs[2 * i + 1]));
+    }
+
+    var sum = 0.0;
+    var src1_i = src1_idx_base + offset * 256;
+    var is: u32 = 0;
+    var m: u32 = 1;
+    // 2 halves of the block (128 elements each)
+    for (var q_b_idx: u32 = 0; q_b_idx < 64; q_b_idx += 32) {
+        // 4 groups (each group has 2 blocks of 16 elements)
+        for (var shift: u32 = 0; shift < 8; shift += 2) {
+            // 2 blocks
+            for (var k: u32 = 0; k < 32; k += 16) {
+                let sc = get_byte(scale_vals[is / 4], is % 4);
+                is++;
+                let dl = d * (f32(sc) - 32.0);
+                for (var l: u32 = 0u; l < 16u; l++) {
+                    let q_idx = q_b_idx + k + l;
+                    let hm_idx = k + l;
+                    let q_byte = get_byte(qs_vals[q_idx / 4], q_idx % 4);
+                    let hmask_byte = get_byte(hmask_vals[hm_idx / 4], hm_idx % 4);
+                    let hm = select(4.0, 0.0, (hmask_byte & m) != 0);
+                    let qs_val = (q_byte >> shift) & 3;
+                    sum += ((f32(qs_val) - hm) * dl) * src1[src1_i];
+                    src1_i++;
+                }
+            }
+            m <<= 1;
+        }
+    }
+    return sum;
+}
+
+#enddecl(Q3_K)
+
+#decl(Q45_K_SCALE_MIN)
+
+fn get_scale_min(is: u32, scales: array<u32, 3>) -> vec2<f32> {
+    if (is < 4) {
+        let sc_byte = get_byte(scales[is / 4], is % 4);
+        let min_byte = get_byte(scales[(is + 4) / 4], is % 4);
+        return vec2(f32(sc_byte & 63), f32(min_byte & 63));
+    } else {
+        let sc_min_lo = get_byte(scales[(is + 4) / 4], (is + 4) % 4);
+        let sc_hi = get_byte(scales[(is - 4) / 4], (is - 4) % 4);
+        let min_hi = get_byte(scales[is / 4], is % 4);
+        let sc = (sc_min_lo & 0xF) | ((sc_hi >> 6) << 4);
+        let m = (sc_min_lo >> 4) | ((min_hi >> 6) << 4);
+        return vec2(f32(sc), f32(m));
+    }
+}
+
+#enddecl(Q45_K_SCALE_MIN)
+
+#decl(Q4_K)
+// 8 blocks of 32 elements each
+struct q4_k {
+    d: f16,
+    dmin: f16,
+    scales: array<u32, 3>,
+    qs: array<u32, 32>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    let m = f32(block.dmin);
+    var sum = 0.0;
+    var src1_i = src1_idx_base + offset * 256;
+    var is: u32 = 0;
+    // 2 blocks each iteration
+    for (var q_b_idx: u32 = 0; q_b_idx < 128; q_b_idx += 32) {
+        for (var shift: u32 = 0; shift < 8; shift += 4) {
+            let scale_min = get_scale_min(is, block.scales);
+            is++;
+            let dl = d * scale_min.x;
+            let ml = m * scale_min.y;
+            for (var l: u32 = 0; l < 32; l++) {
+                let q_idx = q_b_idx + l;
+                let q_byte = get_byte(block.qs[q_idx / 4], q_idx % 4);
+                let qs_val = (q_byte >> shift) & 0xF;
+                sum += (f32(qs_val) * dl - ml) * src1[src1_i];
+                src1_i++;
+            }
+        }
+    }
+    return sum;
+}
+
+#enddecl(Q4_K)
+
+#decl(Q5_K)
+// 8 blocks of 32 elements each
+struct q5_k {
+    d: f16,
+    dmin: f16,
+    scales: array<u32, 3>,
+    qh: array<u32, 8>,
+    qs: array<u32, 32>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    let m = f32(block.dmin);
+    var sum = 0.0;
+    var src1_i = src1_idx_base + offset * 256;
+    var is: u32 = 0;
+    var u: u32 = 1;
+    // 2 blocks each iteration
+    for (var q_b_idx: u32 = 0; q_b_idx < 128; q_b_idx += 32) {
+        for (var shift: u32 = 0; shift < 8; shift += 4) {
+            let scale_min = get_scale_min(is, block.scales);
+            is++;
+            let dl = d * scale_min.x;
+            let ml = m * scale_min.y;
+            for (var l: u32 = 0; l < 32; l++) {
+                let q_idx = q_b_idx + l;
+                let q_byte = get_byte(block.qs[q_idx / 4], q_idx % 4);
+                let qh_byte = get_byte(block.qh[l / 4], l % 4);
+                let qs_val = (q_byte >> shift) & 0xF;
+                let qh_val = select(0.0, 16.0, (qh_byte & u) != 0);
+                sum += ((f32(qs_val) + qh_val) * dl - ml) * src1[src1_i];
+               src1_i++;
+            }
+            u <<= 1;
+        }
+    }
+    return sum;
+}
+
+#enddecl(Q5_K)
+
+#decl(Q6_K)
+// 16 blocks of 16 elements each
+struct q6_k {
+    ql: array<f16, 64>,
+    qh: array<f16, 32>,
+    scales: array<f16, 8>,
+    d: f16
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+
+    // convert arrays of f16 -> u32
+    var ql_vals: array<u32, 32>;
+    for (var i: u32 = 0; i < 32; i++) {
+        ql_vals[i] = bitcast<u32>(vec2(block.ql[2 * i], block.ql[2 * i + 1]));
+    }
+    var qh_vals: array<u32, 16>;
+    for (var i: u32 = 0; i < 16; i++) {
+        qh_vals[i] = bitcast<u32>(vec2(block.qh[2 * i], block.qh[2 * i + 1]));
+    }
+    var scale_vals: array<u32, 4>;
+    for (var i: u32 = 0; i < 4; i++) {
+        scale_vals[i] = bitcast<u32>(vec2(block.scales[2 * i], block.scales[2 * i + 1]));
+    }
+
+    var sum = 0.0;
+    var src1_i = src1_idx_base + offset * 256;
+    var qh_b_idx: u32 = 0;
+    var sc_b_idx: u32 = 0;
+    for (var ql_b_idx: u32 = 0; ql_b_idx < 128; ql_b_idx += 64) {
+        for (var l: u32 = 0; l < 32; l++) {
+            let ql13_b = get_byte(ql_vals[(ql_b_idx + l) / 4], (ql_b_idx + l) % 4);
+            let ql24_b = get_byte(ql_vals[(ql_b_idx + l + 32) / 4], (ql_b_idx + l + 32) % 4);
+            let qh_b = get_byte(qh_vals[(qh_b_idx + l) / 4], (qh_b_idx + l) % 4);
+
+            let q1 = f32((ql13_b & 0xF) | ((qh_b & 3) << 4)) - 32.0;
+            let q2 = f32((ql24_b & 0xF) | (((qh_b >> 2) & 3) << 4)) - 32.0;
+            let q3 = f32((ql13_b >> 4) | (((qh_b >> 4) & 3) << 4)) - 32.0;
+            let q4 = f32((ql24_b >> 4) | (((qh_b >> 6) & 3) << 4)) - 32.0;
+
+            let is = l/16;
+            let is1 = sc_b_idx + is;
+            let sc1 = get_byte_i32(scale_vals[is1 / 4], is1 % 4);
+            let is2 = sc_b_idx + is + 2;
+            let sc2 = get_byte_i32(scale_vals[is2 / 4], is2 % 4);
+            let is3 = sc_b_idx + is + 4;
+            let sc3 = get_byte_i32(scale_vals[is3 / 4], is3 % 4);
+            let is4 = sc_b_idx + is + 6;
+            let sc4 = get_byte_i32(scale_vals[is4 / 4], is4 % 4);
+
+            sum += d * f32(sc1) * q1 * src1[src1_i + l];
+            sum += d * f32(sc2) * q2 * src1[src1_i + l + 32];
+            sum += d * f32(sc3) * q3 * src1[src1_i + l + 64];
+            sum += d * f32(sc4) * q4 * src1[src1_i + l + 96];
+        }
+        src1_i += 128;
+        qh_b_idx += 32;
+        sc_b_idx += 8;
+    }
+    return sum;
+}
+
+#enddecl(Q6_K)
+
+#decl(IQ23_TABLES)
+const kmask_iq2xs : array<u32, 2> = array<u32, 2>(
+    0x08040201u, // 1, 2, 4, 8
+    0x80402010u  // 16, 32, 64, 128
+);
+
+const ksigns_iq2xs: array<u32, 32> = array<u32, 32>(
+    0x03828100,0x87060584,0x8b0a0988,0x0f8e8d0c,
+    0x93121190,0x17969514,0x1b9a9918,0x9f1e1d9c,
+    0xa32221a0,0x27a6a524,0x2baaa928,0xaf2e2dac,
+    0x33b2b130,0xb73635b4,0xbb3a39b8,0x3fbebd3c,
+    0xc34241c0,0x47c6c544,0x4bcac948,0xcf4e4dcc,
+    0x53d2d150,0xd75655d4,0xdb5a59d8,0x5fdedd5c,
+    0x63e2e160,0xe76665e4,0xeb6a69e8,0x6feeed6c,
+    0xf37271f0,0x77f6f574,0x7bfaf978,0xff7e7dfc
+);
+#enddecl(IQ23_TABLES)
+
+#decl(IQ2_XXS)
+
+const iq2xxs_grid = array<u32, 512>(
+    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
+    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x082b0808, 0x08080808,
+    0x082b082b, 0x08080808, 0x082b2b08, 0x08080808, 0x082b2b2b, 0x08080808, 0x19080819, 0x08080808,
+    0x19081908, 0x08080808, 0x19190808, 0x08080808, 0x19192b08, 0x08080808, 0x192b0819, 0x08080808,
+    0x192b1908, 0x08080808, 0x2b080808, 0x08080808, 0x2b08082b, 0x08080808, 0x2b082b2b, 0x08080808,
+    0x2b2b082b, 0x08080808, 0x08080819, 0x08080819, 0x08081908, 0x08080819, 0x08190808, 0x08080819,
+    0x08191919, 0x08080819, 0x19080808, 0x08080819, 0x2b081908, 0x08080819, 0x2b192b08, 0x08080819,
+    0x08080808, 0x0808082b, 0x0808082b, 0x0808082b, 0x082b082b, 0x0808082b, 0x2b08082b, 0x0808082b,
+    0x08080819, 0x08081908, 0x08081908, 0x08081908, 0x08190808, 0x08081908, 0x082b0819, 0x08081908,
+    0x082b1908, 0x08081908, 0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19082b08, 0x08081908,
+    0x192b0808, 0x08081908, 0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b190808, 0x08081908,
+    0x2b2b1908, 0x08081908, 0x08080808, 0x08081919, 0x0808082b, 0x08081919, 0x08082b08, 0x08081919,
+    0x082b0808, 0x08081919, 0x1908192b, 0x08081919, 0x192b2b19, 0x08081919, 0x2b080808, 0x08081919,
+    0x2b190819, 0x08081919, 0x08082b19, 0x0808192b, 0x08190808, 0x0808192b, 0x19080808, 0x0808192b,
+    0x2b081908, 0x0808192b, 0x2b2b1908, 0x0808192b, 0x08080808, 0x08082b08, 0x08081919, 0x08082b08,
+    0x08082b08, 0x08082b08, 0x08191908, 0x08082b08, 0x082b2b08, 0x08082b08, 0x19080819, 0x08082b08,
+    0x19081908, 0x08082b08, 0x19190808, 0x08082b08, 0x1919082b, 0x08082b08, 0x2b082b08, 0x08082b08,
+    0x08081908, 0x08082b19, 0x19080808, 0x08082b19, 0x0808082b, 0x08082b2b, 0x08191908, 0x08082b2b,
+    0x08080819, 0x08190808, 0x08081908, 0x08190808, 0x08190808, 0x08190808, 0x082b0819, 0x08190808,
+    0x19080808, 0x08190808, 0x192b0808, 0x08190808, 0x2b081908, 0x08190808, 0x2b190808, 0x08190808,
+    0x2b191919, 0x08190808, 0x08080808, 0x08190819, 0x08082b08, 0x08190819, 0x082b0808, 0x08190819,
+    0x19190808, 0x08190819, 0x19192b2b, 0x08190819, 0x2b080808, 0x08190819, 0x082b1908, 0x0819082b,
+    0x19081919, 0x0819082b, 0x08080808, 0x08191908, 0x08082b08, 0x08191908, 0x082b0808, 0x08191908,
+    0x082b1919, 0x08191908, 0x19082b19, 0x08191908, 0x2b080808, 0x08191908, 0x08192b08, 0x08191919,
+    0x192b082b, 0x08191919, 0x08080808, 0x0819192b, 0x0819192b, 0x0819192b, 0x08080819, 0x08192b08,
+    0x08081908, 0x08192b08, 0x08190808, 0x08192b08, 0x19080808, 0x08192b08, 0x2b080819, 0x08192b08,
+    0x08080808, 0x08192b19, 0x08081919, 0x08192b19, 0x2b2b0808, 0x08192b19, 0x19190819, 0x08192b2b,
+    0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08082b2b, 0x082b0808, 0x19081908, 0x082b0808,
+    0x192b0819, 0x082b0808, 0x2b080808, 0x082b0808, 0x2b08082b, 0x082b0808, 0x082b2b19, 0x082b0819,
+    0x19082b08, 0x082b0819, 0x08080808, 0x082b082b, 0x0808082b, 0x082b082b, 0x08080819, 0x082b1908,
+    0x08081908, 0x082b1908, 0x08190808, 0x082b1908, 0x19080808, 0x082b1908, 0x1919192b, 0x082b1908,
+    0x08080808, 0x082b1919, 0x19080819, 0x082b1919, 0x192b1908, 0x082b1919, 0x2b190808, 0x082b192b,
+    0x08082b08, 0x082b2b08, 0x082b0808, 0x082b2b08, 0x2b191908, 0x082b2b08, 0x19081908, 0x082b2b2b,
+    0x08080819, 0x19080808, 0x08081908, 0x19080808, 0x08190808, 0x19080808, 0x08192b08, 0x19080808,
+    0x082b0819, 0x19080808, 0x082b1908, 0x19080808, 0x19080808, 0x19080808, 0x19082b08, 0x19080808,
+    0x1919192b, 0x19080808, 0x192b0808, 0x19080808, 0x2b080819, 0x19080808, 0x2b081908, 0x19080808,
+    0x2b190808, 0x19080808, 0x08080808, 0x19080819, 0x082b0808, 0x19080819, 0x192b0819, 0x19080819,
+    0x2b080808, 0x19080819, 0x2b081919, 0x19080819, 0x08080819, 0x1908082b, 0x08190808, 0x1908082b,
+    0x19082b08, 0x1908082b, 0x1919192b, 0x1908082b, 0x192b2b08, 0x1908082b, 0x08080808, 0x19081908,
+    0x08082b08, 0x19081908, 0x082b0808, 0x19081908, 0x2b080808, 0x19081908, 0x2b192b19, 0x19081908,
+    0x0819082b, 0x19081919, 0x082b1908, 0x19081919, 0x08080808, 0x1908192b, 0x08080819, 0x19082b08,
+    0x08081908, 0x19082b08, 0x08190808, 0x19082b08, 0x19080808, 0x19082b08, 0x19081919, 0x19082b08,
+    0x08080808, 0x19082b19, 0x19192b08, 0x19082b19, 0x192b0819, 0x19082b19, 0x2b08082b, 0x19082b19,
+    0x19081919, 0x19082b2b, 0x2b190808, 0x19082b2b, 0x08080808, 0x19190808, 0x08082b08, 0x19190808,
+    0x08190819, 0x19190808, 0x08192b19, 0x19190808, 0x082b0808, 0x19190808, 0x2b080808, 0x19190808,
+    0x2b082b08, 0x19190808, 0x08081908, 0x19190819, 0x1908082b, 0x19190819, 0x2b2b1908, 0x19190819,
+    0x2b190819, 0x1919082b, 0x2b190808, 0x19191908, 0x2b19082b, 0x19191908, 0x08082b2b, 0x19191919,
+    0x08080819, 0x1919192b, 0x19191908, 0x1919192b, 0x08080808, 0x19192b08, 0x08190819, 0x19192b08,
+    0x08192b19, 0x19192b08, 0x192b1908, 0x19192b08, 0x19080808, 0x19192b19, 0x08082b08, 0x19192b2b,
+    0x08081908, 0x192b0808, 0x08190808, 0x192b0808, 0x19080808, 0x192b0808, 0x192b2b08, 0x192b0808,
+    0x08080808, 0x192b0819, 0x19191919, 0x192b0819, 0x08192b08, 0x192b082b, 0x192b0808, 0x192b082b,
+    0x08080808, 0x192b1908, 0x08081919, 0x192b1908, 0x08190808, 0x192b1919, 0x0819082b, 0x192b1919,
+    0x2b081908, 0x192b1919, 0x1908082b, 0x192b2b08, 0x08080808, 0x2b080808, 0x0808082b, 0x2b080808,
+    0x08082b2b, 0x2b080808, 0x19080819, 0x2b080808, 0x2b08082b, 0x2b080808, 0x08081908, 0x2b080819,
+    0x08192b08, 0x2b080819, 0x19080808, 0x2b080819, 0x08190819, 0x2b08082b, 0x08080819, 0x2b081908,
+    0x08081908, 0x2b081908, 0x08190808, 0x2b081908, 0x08191919, 0x2b081908, 0x19080808, 0x2b081908,
+    0x192b0808, 0x2b081908, 0x08080808, 0x2b081919, 0x1908192b, 0x2b081919, 0x2b191908, 0x2b081919,
+    0x08082b19, 0x2b08192b, 0x19080808, 0x2b08192b, 0x192b0808, 0x2b08192b, 0x0808082b, 0x2b082b08,
+    0x08081908, 0x2b082b19, 0x08190819, 0x2b082b2b, 0x08081908, 0x2b190808, 0x08190808, 0x2b190808,
+    0x082b1908, 0x2b190808, 0x19080808, 0x2b190808, 0x2b2b0819, 0x2b190808, 0x0819192b, 0x2b190819,
+    0x2b080808, 0x2b190819, 0x19081919, 0x2b19082b, 0x08080808, 0x2b191908, 0x082b082b, 0x2b191908,
+    0x19081908, 0x2b191908, 0x19190819, 0x2b191919, 0x2b080819, 0x2b192b08, 0x082b0808, 0x2b192b19,
+    0x0808082b, 0x2b2b0808, 0x19190808, 0x2b2b0808, 0x2b081919, 0x2b2b0808, 0x08082b19, 0x2b2b0819,
+    0x08080808, 0x2b2b082b, 0x08192b08, 0x2b2b1908, 0x19190808, 0x2b2b2b08, 0x08081908, 0x2b2b2b19
+);
+
+struct iq2_xxs {
+    d: f16,
+    qs: array<f16, 32>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    var src1_i = src1_idx_base + offset * 256;
+    var sum = 0.0;
+    for (var ib: u32 = 0; ib < 32; ib += 4) {
+        let aux0 = bitcast<u32>(vec2(block.qs[ib], block.qs[ib + 1]));
+        let aux1 = bitcast<u32>(vec2(block.qs[ib + 2], block.qs[ib + 3]));
+        let db = d * (0.5 + f32(aux1 >> 28)) * 0.25;
+        for (var l: u32 = 0; l < 4; l++) {
+            let ig = get_byte(aux0, l) * 8;
+            let is = (aux1 >> (7 * l)) & 127;
+            let signs = get_byte(ksigns_iq2xs[is / 4], is % 4);
+            for (var j: u32 = 0; j < 8; j++) {
+                let g = get_byte(iq2xxs_grid[(ig + j) / 4], (ig + j) % 4);
+                let m = select(1.0, -1.0, (get_byte(kmask_iq2xs[j / 4], j % 4) & signs) != 0);
+                sum += db * f32(g) * m * src1[src1_i];
+                src1_i++;
+            }
+        }
+    }
+    return sum;
+}
+
+#enddecl(IQ2_XXS)
+
+#decl(IQ2_XS)
+const iq2xs_grid = array<u32, 1024>(
+    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
+    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x0819192b, 0x08080808,
+    0x08192b19, 0x08080808, 0x082b0808, 0x08080808, 0x082b082b, 0x08080808, 0x082b1919, 0x08080808,
+    0x082b2b08, 0x08080808, 0x19080819, 0x08080808, 0x19081908, 0x08080808, 0x1908192b, 0x08080808,
+    0x19082b19, 0x08080808, 0x19190808, 0x08080808, 0x1919082b, 0x08080808, 0x19191919, 0x08080808,
+    0x19192b08, 0x08080808, 0x192b0819, 0x08080808, 0x192b1908, 0x08080808, 0x2b080808, 0x08080808,
+    0x2b08082b, 0x08080808, 0x2b081919, 0x08080808, 0x2b082b08, 0x08080808, 0x2b190819, 0x08080808,
+    0x2b191908, 0x08080808, 0x2b192b19, 0x08080808, 0x2b2b0808, 0x08080808, 0x08080819, 0x08080819,
+    0x08081908, 0x08080819, 0x0808192b, 0x08080819, 0x08082b19, 0x08080819, 0x08190808, 0x08080819,
+    0x0819082b, 0x08080819, 0x08191919, 0x08080819, 0x08192b08, 0x08080819, 0x08192b2b, 0x08080819,
+    0x082b0819, 0x08080819, 0x082b1908, 0x08080819, 0x19080808, 0x08080819, 0x1908082b, 0x08080819,
+    0x19081919, 0x08080819, 0x19082b08, 0x08080819, 0x19190819, 0x08080819, 0x19191908, 0x08080819,
+    0x192b0808, 0x08080819, 0x192b2b08, 0x08080819, 0x2b080819, 0x08080819, 0x2b081908, 0x08080819,
+    0x2b190808, 0x08080819, 0x08080808, 0x0808082b, 0x0808082b, 0x0808082b, 0x08081919, 0x0808082b,
+    0x08082b08, 0x0808082b, 0x08190819, 0x0808082b, 0x08191908, 0x0808082b, 0x082b0808, 0x0808082b,
+    0x19080819, 0x0808082b, 0x19081908, 0x0808082b, 0x19190808, 0x0808082b, 0x19191919, 0x0808082b,
+    0x2b080808, 0x0808082b, 0x2b082b2b, 0x0808082b, 0x08080819, 0x08081908, 0x08081908, 0x08081908,
+    0x0808192b, 0x08081908, 0x08082b19, 0x08081908, 0x08190808, 0x08081908, 0x0819082b, 0x08081908,
+    0x08191919, 0x08081908, 0x08192b08, 0x08081908, 0x082b0819, 0x08081908, 0x082b1908, 0x08081908,
+    0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19081919, 0x08081908, 0x19082b08, 0x08081908,
+    0x19190819, 0x08081908, 0x19191908, 0x08081908, 0x1919192b, 0x08081908, 0x192b0808, 0x08081908,
+    0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b190808, 0x08081908, 0x08080808, 0x08081919,
+    0x0808082b, 0x08081919, 0x08081919, 0x08081919, 0x08082b08, 0x08081919, 0x08190819, 0x08081919,
+    0x08191908, 0x08081919, 0x082b0808, 0x08081919, 0x19080819, 0x08081919, 0x19081908, 0x08081919,
+    0x19190808, 0x08081919, 0x192b0819, 0x08081919, 0x2b080808, 0x08081919, 0x08080819, 0x0808192b,
+    0x08081908, 0x0808192b, 0x08190808, 0x0808192b, 0x082b192b, 0x0808192b, 0x19080808, 0x0808192b,
+    0x1908082b, 0x0808192b, 0x2b081908, 0x0808192b, 0x08080808, 0x08082b08, 0x0808082b, 0x08082b08,
+    0x08081919, 0x08082b08, 0x08082b08, 0x08082b08, 0x08082b2b, 0x08082b08, 0x08190819, 0x08082b08,
+    0x08191908, 0x08082b08, 0x082b0808, 0x08082b08, 0x082b1919, 0x08082b08, 0x19080819, 0x08082b08,
+    0x19081908, 0x08082b08, 0x19190808, 0x08082b08, 0x19192b08, 0x08082b08, 0x2b080808, 0x08082b08,
+    0x2b2b0808, 0x08082b08, 0x2b2b2b2b, 0x08082b08, 0x08080819, 0x08082b19, 0x08081908, 0x08082b19,
+    0x08190808, 0x08082b19, 0x19080808, 0x08082b19, 0x2b080819, 0x08082b19, 0x2b082b19, 0x08082b19,
+    0x08080808, 0x08082b2b, 0x082b0808, 0x08082b2b, 0x082b2b08, 0x08082b2b, 0x2b19192b, 0x08082b2b,
+    0x2b2b0808, 0x08082b2b, 0x08080819, 0x08190808, 0x08081908, 0x08190808, 0x0808192b, 0x08190808,
+    0x08082b19, 0x08190808, 0x08190808, 0x08190808, 0x0819082b, 0x08190808, 0x08191919, 0x08190808,
+    0x08192b08, 0x08190808, 0x082b0819, 0x08190808, 0x082b1908, 0x08190808, 0x19080808, 0x08190808,
+    0x1908082b, 0x08190808, 0x19081919, 0x08190808, 0x19082b08, 0x08190808, 0x19190819, 0x08190808,
+    0x19191908, 0x08190808, 0x192b0808, 0x08190808, 0x192b2b2b, 0x08190808, 0x2b080819, 0x08190808,
+    0x2b081908, 0x08190808, 0x2b190808, 0x08190808, 0x08080808, 0x08190819, 0x0808082b, 0x08190819,
+    0x08081919, 0x08190819, 0x08082b08, 0x08190819, 0x08190819, 0x08190819, 0x08191908, 0x08190819,
+    0x082b0808, 0x08190819, 0x19080819, 0x08190819, 0x19081908, 0x08190819, 0x19190808, 0x08190819,
+    0x2b080808, 0x08190819, 0x2b191908, 0x08190819, 0x2b19192b, 0x08190819, 0x08080819, 0x0819082b,
+    0x08081908, 0x0819082b, 0x0808192b, 0x0819082b, 0x08190808, 0x0819082b, 0x19080808, 0x0819082b,
+    0x192b0808, 0x0819082b, 0x08080808, 0x08191908, 0x0808082b, 0x08191908, 0x08081919, 0x08191908,
+    0x08082b08, 0x08191908, 0x08190819, 0x08191908, 0x08191908, 0x08191908, 0x082b0808, 0x08191908,
+    0x19080819, 0x08191908, 0x19081908, 0x08191908, 0x19082b19, 0x08191908, 0x19190808, 0x08191908,
+    0x192b1908, 0x08191908, 0x2b080808, 0x08191908, 0x08080819, 0x08191919, 0x08081908, 0x08191919,
+    0x08190808, 0x08191919, 0x19080808, 0x08191919, 0x08080808, 0x0819192b, 0x08191908, 0x0819192b,
+    0x19082b19, 0x0819192b, 0x08080819, 0x08192b08, 0x08081908, 0x08192b08, 0x08190808, 0x08192b08,
+    0x0819082b, 0x08192b08, 0x19080808, 0x08192b08, 0x19191908, 0x08192b08, 0x2b08192b, 0x08192b08,
+    0x08080808, 0x08192b19, 0x08081919, 0x08192b19, 0x192b192b, 0x08192b19, 0x19190819, 0x08192b2b,
+    0x2b2b2b19, 0x08192b2b, 0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08081919, 0x082b0808,
+    0x08082b08, 0x082b0808, 0x08082b2b, 0x082b0808, 0x08190819, 0x082b0808, 0x08191908, 0x082b0808,
+    0x082b0808, 0x082b0808, 0x19080819, 0x082b0808, 0x19081908, 0x082b0808, 0x19190808, 0x082b0808,
+    0x2b080808, 0x082b0808, 0x2b2b0808, 0x082b0808, 0x08080819, 0x082b0819, 0x08081908, 0x082b0819,
+    0x08190808, 0x082b0819, 0x19080808, 0x082b0819, 0x19082b08, 0x082b0819, 0x192b1919, 0x082b0819,
+    0x08080808, 0x082b082b, 0x082b082b, 0x082b082b, 0x2b080808, 0x082b082b, 0x2b2b2b08, 0x082b082b,
+    0x08080819, 0x082b1908, 0x08081908, 0x082b1908, 0x08190808, 0x082b1908, 0x082b2b19, 0x082b1908,
+    0x19080808, 0x082b1908, 0x08080808, 0x082b1919, 0x19080819, 0x082b1919, 0x1919082b, 0x082b1919,
+    0x2b192b19, 0x082b1919, 0x08080819, 0x082b192b, 0x08192b2b, 0x082b192b, 0x2b2b192b, 0x082b192b,
+    0x08080808, 0x082b2b08, 0x08082b08, 0x082b2b08, 0x08082b2b, 0x082b2b08, 0x082b0808, 0x082b2b08,
+    0x19191919, 0x082b2b08, 0x2b082b08, 0x082b2b08, 0x2b2b082b, 0x082b2b08, 0x192b2b08, 0x082b2b19,
+    0x2b190808, 0x082b2b19, 0x08082b08, 0x082b2b2b, 0x082b0808, 0x082b2b2b, 0x2b08082b, 0x082b2b2b,
+    0x2b082b08, 0x082b2b2b, 0x2b082b2b, 0x082b2b2b, 0x08080819, 0x19080808, 0x08081908, 0x19080808,
+    0x0808192b, 0x19080808, 0x08082b19, 0x19080808, 0x08190808, 0x19080808, 0x0819082b, 0x19080808,
+    0x08191919, 0x19080808, 0x08192b08, 0x19080808, 0x082b0819, 0x19080808, 0x082b1908, 0x19080808,
+    0x19080808, 0x19080808, 0x1908082b, 0x19080808, 0x19081919, 0x19080808, 0x19082b08, 0x19080808,
+    0x19082b2b, 0x19080808, 0x19190819, 0x19080808, 0x19191908, 0x19080808, 0x192b0808, 0x19080808,
+    0x192b1919, 0x19080808, 0x2b080819, 0x19080808, 0x2b081908, 0x19080808, 0x2b190808, 0x19080808,
+    0x08080808, 0x19080819, 0x0808082b, 0x19080819, 0x08081919, 0x19080819, 0x08082b08, 0x19080819,
+    0x08190819, 0x19080819, 0x08191908, 0x19080819, 0x082b0808, 0x19080819, 0x19080819, 0x19080819,
+    0x19081908, 0x19080819, 0x19190808, 0x19080819, 0x2b080808, 0x19080819, 0x2b081919, 0x19080819,
+    0x2b2b082b, 0x19080819, 0x08080819, 0x1908082b, 0x08081908, 0x1908082b, 0x08190808, 0x1908082b,
+    0x0819082b, 0x1908082b, 0x082b2b19, 0x1908082b, 0x19080808, 0x1908082b, 0x08080808, 0x19081908,
+    0x0808082b, 0x19081908, 0x08081919, 0x19081908, 0x08082b08, 0x19081908, 0x08190819, 0x19081908,
+    0x08191908, 0x19081908, 0x08192b19, 0x19081908, 0x082b0808, 0x19081908, 0x19080819, 0x19081908,
+    0x19081908, 0x19081908, 0x19190808, 0x19081908, 0x2b080808, 0x19081908, 0x2b191908, 0x19081908,
+    0x08080819, 0x19081919, 0x08081908, 0x19081919, 0x08190808, 0x19081919, 0x082b1908, 0x19081919,
+    0x19080808, 0x19081919, 0x2b192b2b, 0x19081919, 0x08080808, 0x1908192b, 0x08082b2b, 0x1908192b,
+    0x19081908, 0x1908192b, 0x19190808, 0x1908192b, 0x08080819, 0x19082b08, 0x08081908, 0x19082b08,
+    0x08190808, 0x19082b08, 0x19080808, 0x19082b08, 0x19081919, 0x19082b08, 0x19191908, 0x19082b08,
+    0x192b082b, 0x19082b08, 0x08080808, 0x19082b19, 0x08190819, 0x19082b19, 0x19081908, 0x19082b19,
+    0x19190808, 0x19082b19, 0x192b2b19, 0x19082b19, 0x08081908, 0x19082b2b, 0x08080808, 0x19190808,
+    0x0808082b, 0x19190808, 0x08081919, 0x19190808, 0x08082b08, 0x19190808, 0x08190819, 0x19190808,
+    0x08191908, 0x19190808, 0x082b0808, 0x19190808, 0x082b2b08, 0x19190808, 0x19080819, 0x19190808,
+    0x19081908, 0x19190808, 0x19190808, 0x19190808, 0x2b080808, 0x19190808, 0x08080819, 0x19190819,
+    0x08081908, 0x19190819, 0x08190808, 0x19190819, 0x08191919, 0x19190819, 0x19080808, 0x19190819,
+    0x1908082b, 0x19190819, 0x08080808, 0x1919082b, 0x19081908, 0x1919082b, 0x2b2b2b2b, 0x1919082b,
+    0x08080819, 0x19191908, 0x08081908, 0x19191908, 0x08190808, 0x19191908, 0x082b0819, 0x19191908,
+    0x19080808, 0x19191908, 0x192b0808, 0x19191908, 0x2b080819, 0x19191908, 0x2b2b0819, 0x19191908,
+    0x08080808, 0x19191919, 0x08082b08, 0x19191919, 0x2b080808, 0x19191919, 0x2b082b08, 0x19191919,
+    0x082b0819, 0x1919192b, 0x192b2b08, 0x1919192b, 0x2b2b0819, 0x1919192b, 0x08080808, 0x19192b08,
+    0x08191908, 0x19192b08, 0x19080819, 0x19192b08, 0x19190808, 0x19192b08, 0x2b192b19, 0x19192b08,
+    0x08192b2b, 0x19192b19, 0x19080808, 0x19192b19, 0x1908082b, 0x19192b19, 0x2b081919, 0x19192b2b,
+    0x08080819, 0x192b0808, 0x08081908, 0x192b0808, 0x08190808, 0x192b0808, 0x19080808, 0x192b0808,
+    0x19191908, 0x192b0808, 0x192b082b, 0x192b0808, 0x2b08192b, 0x192b0808, 0x2b2b2b19, 0x192b0808,
+    0x08080808, 0x192b0819, 0x082b1908, 0x192b082b, 0x19082b2b, 0x192b082b, 0x2b19082b, 0x192b082b,
+    0x08080808, 0x192b1908, 0x0819192b, 0x192b1908, 0x08190808, 0x192b1919, 0x19080808, 0x192b1919,
+    0x19081919, 0x192b1919, 0x2b2b1908, 0x192b1919, 0x08080819, 0x192b2b08, 0x192b2b2b, 0x192b2b08,
+    0x082b1919, 0x192b2b19, 0x0808192b, 0x192b2b2b, 0x19191908, 0x192b2b2b, 0x192b082b, 0x192b2b2b,
+    0x08080808, 0x2b080808, 0x0808082b, 0x2b080808, 0x08081919, 0x2b080808, 0x08082b08, 0x2b080808,
+    0x08190819, 0x2b080808, 0x08191908, 0x2b080808, 0x082b0808, 0x2b080808, 0x082b2b2b, 0x2b080808,
+    0x19080819, 0x2b080808, 0x19081908, 0x2b080808, 0x19190808, 0x2b080808, 0x2b080808, 0x2b080808,
+    0x2b08082b, 0x2b080808, 0x2b2b2b08, 0x2b080808, 0x2b2b2b2b, 0x2b080808, 0x08080819, 0x2b080819,
+    0x08081908, 0x2b080819, 0x0808192b, 0x2b080819, 0x08190808, 0x2b080819, 0x19080808, 0x2b080819,
+    0x19190819, 0x2b080819, 0x19192b19, 0x2b080819, 0x08080808, 0x2b08082b, 0x082b0808, 0x2b08082b,
+    0x2b080808, 0x2b08082b, 0x2b08082b, 0x2b08082b, 0x2b2b0808, 0x2b08082b, 0x2b2b2b08, 0x2b08082b,
+    0x08080819, 0x2b081908, 0x08081908, 0x2b081908, 0x08190808, 0x2b081908, 0x0819082b, 0x2b081908,
+    0x08191919, 0x2b081908, 0x19080808, 0x2b081908, 0x192b0808, 0x2b081908, 0x2b082b19, 0x2b081908,
+    0x08080808, 0x2b081919, 0x19081908, 0x2b081919, 0x2b2b1919, 0x2b081919, 0x08192b08, 0x2b08192b,
+    0x192b2b2b, 0x2b08192b, 0x08080808, 0x2b082b08, 0x08082b08, 0x2b082b08, 0x082b1919, 0x2b082b08,
+    0x19192b2b, 0x2b082b08, 0x2b080808, 0x2b082b08, 0x2b08082b, 0x2b082b08, 0x2b2b2b08, 0x2b082b08,
+    0x0808192b, 0x2b082b19, 0x082b082b, 0x2b082b2b, 0x2b080808, 0x2b082b2b, 0x2b082b08, 0x2b082b2b,
+    0x2b19192b, 0x2b082b2b, 0x2b2b2b08, 0x2b082b2b, 0x08080819, 0x2b190808, 0x08081908, 0x2b190808,
+    0x08190808, 0x2b190808, 0x19080808, 0x2b190808, 0x1919192b, 0x2b190808, 0x2b081908, 0x2b190808,
+    0x08080808, 0x2b190819, 0x082b082b, 0x2b190819, 0x192b1908, 0x2b190819, 0x1919192b, 0x2b19082b,
+    0x2b082b19, 0x2b19082b, 0x08080808, 0x2b191908, 0x08081919, 0x2b191908, 0x19081908, 0x2b191908,
+    0x19190808, 0x2b191908, 0x19192b08, 0x2b191908, 0x082b2b19, 0x2b191919, 0x2b190808, 0x2b191919,
+    0x2b19082b, 0x2b191919, 0x19080819, 0x2b19192b, 0x19190819, 0x2b192b08, 0x2b2b192b, 0x2b192b08,
+    0x19082b19, 0x2b192b19, 0x08191919, 0x2b192b2b, 0x192b0808, 0x2b192b2b, 0x08080808, 0x2b2b0808,
+    0x0808082b, 0x2b2b0808, 0x08082b08, 0x2b2b0808, 0x08082b2b, 0x2b2b0808, 0x082b0808, 0x2b2b0808,
+    0x082b2b2b, 0x2b2b0808, 0x2b2b0808, 0x2b2b0808, 0x19190819, 0x2b2b0819, 0x19192b19, 0x2b2b0819,
+    0x2b2b192b, 0x2b2b0819, 0x08080808, 0x2b2b082b, 0x0808082b, 0x2b2b082b, 0x08082b08, 0x2b2b082b,
+    0x082b2b2b, 0x2b2b082b, 0x2b080808, 0x2b2b082b, 0x2b2b0808, 0x2b2b082b, 0x19080808, 0x2b2b1908,
+    0x2b191919, 0x2b2b1908, 0x192b1919, 0x2b2b192b, 0x2b192b08, 0x2b2b192b, 0x08082b2b, 0x2b2b2b08,
+    0x082b0808, 0x2b2b2b08, 0x082b082b, 0x2b2b2b08, 0x082b2b08, 0x2b2b2b08, 0x2b2b0808, 0x2b2b2b08,
+    0x2b2b2b08, 0x2b2b2b08, 0x08081908, 0x2b2b2b19, 0x2b081908, 0x2b2b2b19, 0x2b08192b, 0x2b2b2b19,
+    0x082b2b08, 0x2b2b2b2b, 0x082b2b2b, 0x2b2b2b2b, 0x2b190819, 0x2b2b2b2b, 0x2b2b2b2b, 0x2b2b2b2b
+);
+
+struct iq2_xs {
+    d: f16,
+    qs: array<f16, 32>,
+    scales: array<f16, 4>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    var src1_i = src1_idx_base + offset * 256;
+    var scale_vals = array<u32, 2>(
+        bitcast<u32>(vec2(block.scales[0], block.scales[1])),
+        bitcast<u32>(vec2(block.scales[2], block.scales[3]))
+    );
+    var sum = 0.0;
+    for (var ib: u32 = 0; ib < 32; ib += 4) {
+        let s = get_byte(scale_vals[ib / 16], (ib % 16) / 4);
+        let db = array<f32, 2>(
+            d * (0.5 + f32(s & 0xF)) * 0.25,
+            d * (0.5 + f32(s >> 4)) * 0.25
+        );
+        for (var l: u32 = 0; l < 4; l++) {
+            let qs_val = bitcast<u32>(vec2(block.qs[ib + l], 0.0));
+            let ig = (qs_val & 511) * 8;
+            let is = qs_val >> 9;
+            let signs = get_byte(ksigns_iq2xs[is / 4], is % 4);
+            let dl = db[l/2];
+            for (var j: u32 = 0; j < 8; j++) {
+                let g = get_byte(iq2xs_grid[(ig + j) / 4], (ig + j) % 4);
+                let m = select(1.0, -1.0, (get_byte(kmask_iq2xs[j / 4], j % 4) & signs) != 0);
+                sum += dl * f32(g) * m * src1[src1_i];
+                src1_i++;
+            }
+        }
+    }
+    return sum;
+}
+
+#enddecl(IQ2_XS)
+
+#decl(IQ2_S)
+
+const iq2s_grid = array<u32, 2048>(
+    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
+    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x0819192b, 0x08080808,
+    0x08192b19, 0x08080808, 0x082b0808, 0x08080808, 0x082b082b, 0x08080808, 0x082b1919, 0x08080808,
+    0x082b2b08, 0x08080808, 0x19080819, 0x08080808, 0x19081908, 0x08080808, 0x1908192b, 0x08080808,
+    0x19082b19, 0x08080808, 0x19190808, 0x08080808, 0x1919082b, 0x08080808, 0x19191919, 0x08080808,
+    0x19192b08, 0x08080808, 0x192b0819, 0x08080808, 0x192b1908, 0x08080808, 0x192b192b, 0x08080808,
+    0x192b2b19, 0x08080808, 0x2b080808, 0x08080808, 0x2b08082b, 0x08080808, 0x2b081919, 0x08080808,
+    0x2b082b08, 0x08080808, 0x2b190819, 0x08080808, 0x2b191908, 0x08080808, 0x2b2b0808, 0x08080808,
+    0x2b2b1919, 0x08080808, 0x2b2b2b2b, 0x08080808, 0x08080819, 0x08080819, 0x08081908, 0x08080819,
+    0x0808192b, 0x08080819, 0x08082b19, 0x08080819, 0x08190808, 0x08080819, 0x0819082b, 0x08080819,
+    0x08191919, 0x08080819, 0x08192b08, 0x08080819, 0x082b0819, 0x08080819, 0x082b1908, 0x08080819,
+    0x19080808, 0x08080819, 0x1908082b, 0x08080819, 0x19081919, 0x08080819, 0x19082b08, 0x08080819,
+    0x19190819, 0x08080819, 0x19191908, 0x08080819, 0x1919192b, 0x08080819, 0x19192b19, 0x08080819,
+    0x192b0808, 0x08080819, 0x192b1919, 0x08080819, 0x192b2b08, 0x08080819, 0x2b080819, 0x08080819,
+    0x2b081908, 0x08080819, 0x2b190808, 0x08080819, 0x2b19082b, 0x08080819, 0x2b191919, 0x08080819,
+    0x2b2b0819, 0x08080819, 0x2b2b1908, 0x08080819, 0x08080808, 0x0808082b, 0x0808082b, 0x0808082b,
+    0x08081919, 0x0808082b, 0x08082b08, 0x0808082b, 0x08190819, 0x0808082b, 0x08191908, 0x0808082b,
+    0x082b0808, 0x0808082b, 0x082b2b2b, 0x0808082b, 0x19080819, 0x0808082b, 0x19081908, 0x0808082b,
+    0x1908192b, 0x0808082b, 0x19082b19, 0x0808082b, 0x19190808, 0x0808082b, 0x19191919, 0x0808082b,
+    0x2b080808, 0x0808082b, 0x2b081919, 0x0808082b, 0x2b082b2b, 0x0808082b, 0x2b191908, 0x0808082b,
+    0x2b2b082b, 0x0808082b, 0x08080819, 0x08081908, 0x08081908, 0x08081908, 0x0808192b, 0x08081908,
+    0x08082b19, 0x08081908, 0x08190808, 0x08081908, 0x0819082b, 0x08081908, 0x08191919, 0x08081908,
+    0x08192b08, 0x08081908, 0x082b0819, 0x08081908, 0x082b1908, 0x08081908, 0x082b192b, 0x08081908,
+    0x082b2b19, 0x08081908, 0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19081919, 0x08081908,
+    0x19082b08, 0x08081908, 0x19082b2b, 0x08081908, 0x19190819, 0x08081908, 0x19191908, 0x08081908,
+    0x1919192b, 0x08081908, 0x19192b19, 0x08081908, 0x192b0808, 0x08081908, 0x192b082b, 0x08081908,
+    0x192b1919, 0x08081908, 0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b08192b, 0x08081908,
+    0x2b082b19, 0x08081908, 0x2b190808, 0x08081908, 0x2b191919, 0x08081908, 0x2b192b08, 0x08081908,
+    0x2b2b0819, 0x08081908, 0x2b2b1908, 0x08081908, 0x08080808, 0x08081919, 0x0808082b, 0x08081919,
+    0x08081919, 0x08081919, 0x08082b08, 0x08081919, 0x08082b2b, 0x08081919, 0x08190819, 0x08081919,
+    0x08191908, 0x08081919, 0x0819192b, 0x08081919, 0x08192b19, 0x08081919, 0x082b0808, 0x08081919,
+    0x082b1919, 0x08081919, 0x082b2b08, 0x08081919, 0x19080819, 0x08081919, 0x19081908, 0x08081919,
+    0x1908192b, 0x08081919, 0x19082b19, 0x08081919, 0x19190808, 0x08081919, 0x1919082b, 0x08081919,
+    0x19191919, 0x08081919, 0x19192b08, 0x08081919, 0x192b0819, 0x08081919, 0x192b1908, 0x08081919,
+    0x2b080808, 0x08081919, 0x2b08082b, 0x08081919, 0x2b081919, 0x08081919, 0x2b082b08, 0x08081919,
+    0x2b190819, 0x08081919, 0x2b191908, 0x08081919, 0x2b2b0808, 0x08081919, 0x08080819, 0x0808192b,
+    0x08081908, 0x0808192b, 0x0808192b, 0x0808192b, 0x08082b19, 0x0808192b, 0x08190808, 0x0808192b,
+    0x08191919, 0x0808192b, 0x19080808, 0x0808192b, 0x19081919, 0x0808192b, 0x19082b08, 0x0808192b,
+    0x19190819, 0x0808192b, 0x19191908, 0x0808192b, 0x192b0808, 0x0808192b, 0x2b080819, 0x0808192b,
+    0x2b081908, 0x0808192b, 0x2b190808, 0x0808192b, 0x08080808, 0x08082b08, 0x0808082b, 0x08082b08,
+    0x08081919, 0x08082b08, 0x08082b08, 0x08082b08, 0x08190819, 0x08082b08, 0x08191908, 0x08082b08,
+    0x0819192b, 0x08082b08, 0x08192b19, 0x08082b08, 0x082b0808, 0x08082b08, 0x082b1919, 0x08082b08,
+    0x082b2b2b, 0x08082b08, 0x19080819, 0x08082b08, 0x19081908, 0x08082b08, 0x1908192b, 0x08082b08,
+    0x19082b19, 0x08082b08, 0x19190808, 0x08082b08, 0x1919082b, 0x08082b08, 0x19191919, 0x08082b08,
+    0x19192b08, 0x08082b08, 0x192b0819, 0x08082b08, 0x192b1908, 0x08082b08, 0x2b080808, 0x08082b08,
+    0x2b081919, 0x08082b08, 0x2b191908, 0x08082b08, 0x2b2b2b2b, 0x08082b08, 0x08080819, 0x08082b19,
+    0x08081908, 0x08082b19, 0x08190808, 0x08082b19, 0x0819082b, 0x08082b19, 0x08191919, 0x08082b19,
+    0x08192b08, 0x08082b19, 0x082b0819, 0x08082b19, 0x19080808, 0x08082b19, 0x19081919, 0x08082b19,
+    0x19082b08, 0x08082b19, 0x19190819, 0x08082b19, 0x19191908, 0x08082b19, 0x192b0808, 0x08082b19,
+    0x2b080819, 0x08082b19, 0x2b190808, 0x08082b19, 0x08080808, 0x08082b2b, 0x08190819, 0x08082b2b,
+    0x08191908, 0x08082b2b, 0x082b082b, 0x08082b2b, 0x082b2b08, 0x08082b2b, 0x082b2b2b, 0x08082b2b,
+    0x19190808, 0x08082b2b, 0x2b192b19, 0x08082b2b, 0x08080819, 0x08190808, 0x08081908, 0x08190808,
+    0x0808192b, 0x08190808, 0x08082b19, 0x08190808, 0x08190808, 0x08190808, 0x0819082b, 0x08190808,
+    0x08191919, 0x08190808, 0x08192b08, 0x08190808, 0x082b0819, 0x08190808, 0x082b1908, 0x08190808,
+    0x082b192b, 0x08190808, 0x19080808, 0x08190808, 0x1908082b, 0x08190808, 0x19081919, 0x08190808,
+    0x19082b08, 0x08190808, 0x19190819, 0x08190808, 0x19191908, 0x08190808, 0x1919192b, 0x08190808,
+    0x19192b19, 0x08190808, 0x192b0808, 0x08190808, 0x192b082b, 0x08190808, 0x192b1919, 0x08190808,
+    0x192b2b08, 0x08190808, 0x2b080819, 0x08190808, 0x2b081908, 0x08190808, 0x2b08192b, 0x08190808,
+    0x2b190808, 0x08190808, 0x2b191919, 0x08190808, 0x2b192b08, 0x08190808, 0x2b2b0819, 0x08190808,
+    0x2b2b1908, 0x08190808, 0x08080808, 0x08190819, 0x0808082b, 0x08190819, 0x08081919, 0x08190819,
+    0x08082b08, 0x08190819, 0x08082b2b, 0x08190819, 0x08190819, 0x08190819, 0x08191908, 0x08190819,
+    0x0819192b, 0x08190819, 0x08192b19, 0x08190819, 0x082b0808, 0x08190819, 0x082b082b, 0x08190819,
+    0x082b1919, 0x08190819, 0x082b2b08, 0x08190819, 0x19080819, 0x08190819, 0x19081908, 0x08190819,
+    0x1908192b, 0x08190819, 0x19082b19, 0x08190819, 0x19190808, 0x08190819, 0x1919082b, 0x08190819,
+    0x19191919, 0x08190819, 0x19192b08, 0x08190819, 0x192b0819, 0x08190819, 0x192b1908, 0x08190819,
+    0x2b080808, 0x08190819, 0x2b08082b, 0x08190819, 0x2b081919, 0x08190819, 0x2b082b08, 0x08190819,
+    0x2b190819, 0x08190819, 0x2b191908, 0x08190819, 0x08080819, 0x0819082b, 0x08081908, 0x0819082b,
+    0x08082b19, 0x0819082b, 0x08190808, 0x0819082b, 0x08191919, 0x0819082b, 0x082b0819, 0x0819082b,
+    0x082b1908, 0x0819082b, 0x19080808, 0x0819082b, 0x19081919, 0x0819082b, 0x19190819, 0x0819082b,
+    0x19191908, 0x0819082b, 0x2b080819, 0x0819082b, 0x2b081908, 0x0819082b, 0x2b190808, 0x0819082b,
+    0x08080808, 0x08191908, 0x0808082b, 0x08191908, 0x08081919, 0x08191908, 0x08082b08, 0x08191908,
+    0x08190819, 0x08191908, 0x08191908, 0x08191908, 0x0819192b, 0x08191908, 0x08192b19, 0x08191908,
+    0x082b0808, 0x08191908, 0x082b1919, 0x08191908, 0x082b2b08, 0x08191908, 0x19080819, 0x08191908,
+    0x19081908, 0x08191908, 0x1908192b, 0x08191908, 0x19082b19, 0x08191908, 0x19190808, 0x08191908,
+    0x1919082b, 0x08191908, 0x19191919, 0x08191908, 0x19192b08, 0x08191908, 0x192b0819, 0x08191908,
+    0x192b1908, 0x08191908, 0x2b080808, 0x08191908, 0x2b08082b, 0x08191908, 0x2b081919, 0x08191908,
+    0x2b082b08, 0x08191908, 0x2b190819, 0x08191908, 0x2b191908, 0x08191908, 0x2b2b0808, 0x08191908,
+    0x08080819, 0x08191919, 0x08081908, 0x08191919, 0x0808192b, 0x08191919, 0x08082b19, 0x08191919,
+    0x08190808, 0x08191919, 0x0819082b, 0x08191919, 0x08191919, 0x08191919, 0x08192b08, 0x08191919,
+    0x082b0819, 0x08191919, 0x082b1908, 0x08191919, 0x19080808, 0x08191919, 0x1908082b, 0x08191919,
+    0x19081919, 0x08191919, 0x19082b08, 0x08191919, 0x19190819, 0x08191919, 0x19191908, 0x08191919,
+    0x192b0808, 0x08191919, 0x2b080819, 0x08191919, 0x2b081908, 0x08191919, 0x2b190808, 0x08191919,
+    0x08080808, 0x0819192b, 0x08081919, 0x0819192b, 0x08082b08, 0x0819192b, 0x08190819, 0x0819192b,
+    0x08191908, 0x0819192b, 0x082b0808, 0x0819192b, 0x19080819, 0x0819192b, 0x19081908, 0x0819192b,
+    0x19190808, 0x0819192b, 0x2b080808, 0x0819192b, 0x2b2b2b2b, 0x0819192b, 0x08080819, 0x08192b08,
+    0x08081908, 0x08192b08, 0x0808192b, 0x08192b08, 0x08082b19, 0x08192b08, 0x08190808, 0x08192b08,
+    0x08191919, 0x08192b08, 0x08192b08, 0x08192b08, 0x082b0819, 0x08192b08, 0x19080808, 0x08192b08,
+    0x1908082b, 0x08192b08, 0x19081919, 0x08192b08, 0x19082b08, 0x08192b08, 0x19190819, 0x08192b08,
+    0x19191908, 0x08192b08, 0x192b0808, 0x08192b08, 0x2b080819, 0x08192b08, 0x2b081908, 0x08192b08,
+    0x08080808, 0x08192b19, 0x0808082b, 0x08192b19, 0x08081919, 0x08192b19, 0x08082b08, 0x08192b19,
+    0x08190819, 0x08192b19, 0x08191908, 0x08192b19, 0x082b0808, 0x08192b19, 0x19080819, 0x08192b19,
+    0x19081908, 0x08192b19, 0x19190808, 0x08192b19, 0x192b2b19, 0x08192b19, 0x2b2b082b, 0x08192b19,
+    0x08081908, 0x08192b2b, 0x08190808, 0x08192b2b, 0x19080808, 0x08192b2b, 0x1919192b, 0x08192b2b,
+    0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08081919, 0x082b0808, 0x08082b08, 0x082b0808,
+    0x08190819, 0x082b0808, 0x08191908, 0x082b0808, 0x0819192b, 0x082b0808, 0x08192b19, 0x082b0808,
+    0x082b0808, 0x082b0808, 0x082b1919, 0x082b0808, 0x082b2b2b, 0x082b0808, 0x19080819, 0x082b0808,
+    0x19081908, 0x082b0808, 0x19190808, 0x082b0808, 0x1919082b, 0x082b0808, 0x19191919, 0x082b0808,
+    0x192b1908, 0x082b0808, 0x2b080808, 0x082b0808, 0x2b082b2b, 0x082b0808, 0x2b191908, 0x082b0808,
+    0x2b2b2b2b, 0x082b0808, 0x08080819, 0x082b0819, 0x08081908, 0x082b0819, 0x08190808, 0x082b0819,
+    0x0819082b, 0x082b0819, 0x08191919, 0x082b0819, 0x082b0819, 0x082b0819, 0x19080808, 0x082b0819,
+    0x1908082b, 0x082b0819, 0x19081919, 0x082b0819, 0x19190819, 0x082b0819, 0x19191908, 0x082b0819,
+    0x192b0808, 0x082b0819, 0x2b080819, 0x082b0819, 0x2b081908, 0x082b0819, 0x2b190808, 0x082b0819,
+    0x08080808, 0x082b082b, 0x08082b2b, 0x082b082b, 0x082b082b, 0x082b082b, 0x082b2b08, 0x082b082b,
+    0x082b2b2b, 0x082b082b, 0x19081908, 0x082b082b, 0x19190808, 0x082b082b, 0x2b082b08, 0x082b082b,
+    0x2b082b2b, 0x082b082b, 0x2b2b2b08, 0x082b082b, 0x08080819, 0x082b1908, 0x08081908, 0x082b1908,
+    0x0808192b, 0x082b1908, 0x08082b19, 0x082b1908, 0x08190808, 0x082b1908, 0x08191919, 0x082b1908,
+    0x08192b08, 0x082b1908, 0x082b0819, 0x082b1908, 0x082b1908, 0x082b1908, 0x19080808, 0x082b1908,
+    0x1908082b, 0x082b1908, 0x19081919, 0x082b1908, 0x19082b08, 0x082b1908, 0x19190819, 0x082b1908,
+    0x19191908, 0x082b1908, 0x192b0808, 0x082b1908, 0x2b080819, 0x082b1908, 0x2b081908, 0x082b1908,
+    0x2b190808, 0x082b1908, 0x08080808, 0x082b1919, 0x08081919, 0x082b1919, 0x08082b08, 0x082b1919,
+    0x08190819, 0x082b1919, 0x08191908, 0x082b1919, 0x082b0808, 0x082b1919, 0x19080819, 0x082b1919,
+    0x19081908, 0x082b1919, 0x19190808, 0x082b1919, 0x192b192b, 0x082b1919, 0x2b080808, 0x082b1919,
+    0x08080819, 0x082b192b, 0x08081908, 0x082b192b, 0x08190808, 0x082b192b, 0x19080808, 0x082b192b,
+    0x19192b19, 0x082b192b, 0x08080808, 0x082b2b08, 0x08081919, 0x082b2b08, 0x08190819, 0x082b2b08,
+    0x08191908, 0x082b2b08, 0x19080819, 0x082b2b08, 0x19081908, 0x082b2b08, 0x19190808, 0x082b2b08,
+    0x2b082b2b, 0x082b2b08, 0x2b2b2b2b, 0x082b2b08, 0x08080819, 0x082b2b19, 0x08081908, 0x082b2b19,
+    0x08190808, 0x082b2b19, 0x2b191919, 0x082b2b19, 0x08082b2b, 0x082b2b2b, 0x082b082b, 0x082b2b2b,
+    0x192b1908, 0x082b2b2b, 0x2b082b08, 0x082b2b2b, 0x2b082b2b, 0x082b2b2b, 0x08080819, 0x19080808,
+    0x08081908, 0x19080808, 0x0808192b, 0x19080808, 0x08082b19, 0x19080808, 0x08190808, 0x19080808,
+    0x0819082b, 0x19080808, 0x08191919, 0x19080808, 0x08192b08, 0x19080808, 0x08192b2b, 0x19080808,
+    0x082b0819, 0x19080808, 0x082b1908, 0x19080808, 0x082b192b, 0x19080808, 0x19080808, 0x19080808,
+    0x1908082b, 0x19080808, 0x19081919, 0x19080808, 0x19082b08, 0x19080808, 0x19082b2b, 0x19080808,
+    0x19190819, 0x19080808, 0x19191908, 0x19080808, 0x1919192b, 0x19080808, 0x19192b19, 0x19080808,
+    0x192b0808, 0x19080808, 0x192b082b, 0x19080808, 0x192b1919, 0x19080808, 0x2b080819, 0x19080808,
+    0x2b081908, 0x19080808, 0x2b190808, 0x19080808, 0x2b191919, 0x19080808, 0x2b192b08, 0x19080808,
+    0x2b2b0819, 0x19080808, 0x2b2b1908, 0x19080808, 0x08080808, 0x19080819, 0x0808082b, 0x19080819,
+    0x08081919, 0x19080819, 0x08082b08, 0x19080819, 0x08190819, 0x19080819, 0x08191908, 0x19080819,
+    0x0819192b, 0x19080819, 0x08192b19, 0x19080819, 0x082b0808, 0x19080819, 0x082b082b, 0x19080819,
+    0x082b1919, 0x19080819, 0x19080819, 0x19080819, 0x19081908, 0x19080819, 0x1908192b, 0x19080819,
+    0x19082b19, 0x19080819, 0x19190808, 0x19080819, 0x1919082b, 0x19080819, 0x19191919, 0x19080819,
+    0x19192b08, 0x19080819, 0x192b0819, 0x19080819, 0x192b1908, 0x19080819, 0x2b080808, 0x19080819,
+    0x2b08082b, 0x19080819, 0x2b081919, 0x19080819, 0x2b082b08, 0x19080819, 0x2b190819, 0x19080819,
+    0x2b191908, 0x19080819, 0x2b2b0808, 0x19080819, 0x08080819, 0x1908082b, 0x08081908, 0x1908082b,
+    0x08190808, 0x1908082b, 0x0819082b, 0x1908082b, 0x08191919, 0x1908082b, 0x08192b08, 0x1908082b,
+    0x082b1908, 0x1908082b, 0x19080808, 0x1908082b, 0x19081919, 0x1908082b, 0x19082b08, 0x1908082b,
+    0x19190819, 0x1908082b, 0x19191908, 0x1908082b, 0x192b0808, 0x1908082b, 0x2b080819, 0x1908082b,
+    0x2b081908, 0x1908082b, 0x08080808, 0x19081908, 0x0808082b, 0x19081908, 0x08081919, 0x19081908,
+    0x08082b08, 0x19081908, 0x08082b2b, 0x19081908, 0x08190819, 0x19081908, 0x08191908, 0x19081908,
+    0x0819192b, 0x19081908, 0x08192b19, 0x19081908, 0x082b0808, 0x19081908, 0x082b082b, 0x19081908,
+    0x082b1919, 0x19081908, 0x082b2b08, 0x19081908, 0x19080819, 0x19081908, 0x19081908, 0x19081908,
+    0x1908192b, 0x19081908, 0x19082b19, 0x19081908, 0x19190808, 0x19081908, 0x1919082b, 0x19081908,
+    0x19191919, 0x19081908, 0x19192b08, 0x19081908, 0x192b0819, 0x19081908, 0x192b1908, 0x19081908,
+    0x2b080808, 0x19081908, 0x2b08082b, 0x19081908, 0x2b081919, 0x19081908, 0x2b082b08, 0x19081908,
+    0x2b190819, 0x19081908, 0x2b191908, 0x19081908, 0x2b2b0808, 0x19081908, 0x08080819, 0x19081919,
+    0x08081908, 0x19081919, 0x0808192b, 0x19081919, 0x08082b19, 0x19081919, 0x08190808, 0x19081919,
+    0x0819082b, 0x19081919, 0x08191919, 0x19081919, 0x08192b08, 0x19081919, 0x082b0819, 0x19081919,
+    0x082b1908, 0x19081919, 0x19080808, 0x19081919, 0x1908082b, 0x19081919, 0x19081919, 0x19081919,
+    0x19082b08, 0x19081919, 0x19190819, 0x19081919, 0x19191908, 0x19081919, 0x192b0808, 0x19081919,
+    0x192b2b2b, 0x19081919, 0x2b080819, 0x19081919, 0x2b081908, 0x19081919, 0x2b190808, 0x19081919,
+    0x08080808, 0x1908192b, 0x0808082b, 0x1908192b, 0x08081919, 0x1908192b, 0x08082b08, 0x1908192b,
+    0x08190819, 0x1908192b, 0x08191908, 0x1908192b, 0x082b0808, 0x1908192b, 0x19080819, 0x1908192b,
+    0x19081908, 0x1908192b, 0x19190808, 0x1908192b, 0x2b080808, 0x1908192b, 0x2b2b1919, 0x1908192b,
+    0x08080819, 0x19082b08, 0x08081908, 0x19082b08, 0x08082b19, 0x19082b08, 0x08190808, 0x19082b08,
+    0x0819082b, 0x19082b08, 0x08191919, 0x19082b08, 0x08192b08, 0x19082b08, 0x082b0819, 0x19082b08,
+    0x082b1908, 0x19082b08, 0x19080808, 0x19082b08, 0x1908082b, 0x19082b08, 0x19081919, 0x19082b08,
+    0x19082b08, 0x19082b08, 0x19190819, 0x19082b08, 0x19191908, 0x19082b08, 0x192b0808, 0x19082b08,
+    0x2b081908, 0x19082b08, 0x2b190808, 0x19082b08, 0x08080808, 0x19082b19, 0x0808082b, 0x19082b19,
+    0x08081919, 0x19082b19, 0x08082b08, 0x19082b19, 0x08190819, 0x19082b19, 0x08191908, 0x19082b19,
+    0x082b0808, 0x19082b19, 0x19080819, 0x19082b19, 0x19081908, 0x19082b19, 0x19190808, 0x19082b19,
+    0x2b080808, 0x19082b19, 0x2b19192b, 0x19082b19, 0x08080819, 0x19082b2b, 0x08081908, 0x19082b2b,
+    0x08190808, 0x19082b2b, 0x19080808, 0x19082b2b, 0x08080808, 0x19190808, 0x0808082b, 0x19190808,
+    0x08081919, 0x19190808, 0x08082b08, 0x19190808, 0x08190819, 0x19190808, 0x08191908, 0x19190808,
+    0x0819192b, 0x19190808, 0x08192b19, 0x19190808, 0x082b0808, 0x19190808, 0x082b082b, 0x19190808,
+    0x082b1919, 0x19190808, 0x082b2b08, 0x19190808, 0x19080819, 0x19190808, 0x19081908, 0x19190808,
+    0x1908192b, 0x19190808, 0x19082b19, 0x19190808, 0x19190808, 0x19190808, 0x1919082b, 0x19190808,
+    0x19191919, 0x19190808, 0x19192b08, 0x19190808, 0x192b0819, 0x19190808, 0x192b1908, 0x19190808,
+    0x2b080808, 0x19190808, 0x2b08082b, 0x19190808, 0x2b081919, 0x19190808, 0x2b082b08, 0x19190808,
+    0x2b190819, 0x19190808, 0x2b191908, 0x19190808, 0x08080819, 0x19190819, 0x08081908, 0x19190819,
+    0x0808192b, 0x19190819, 0x08082b19, 0x19190819, 0x08190808, 0x19190819, 0x0819082b, 0x19190819,
+    0x08191919, 0x19190819, 0x08192b08, 0x19190819, 0x082b0819, 0x19190819, 0x082b1908, 0x19190819,
+    0x19080808, 0x19190819, 0x1908082b, 0x19190819, 0x19081919, 0x19190819, 0x19082b08, 0x19190819,
+    0x19190819, 0x19190819, 0x19191908, 0x19190819, 0x192b0808, 0x19190819, 0x2b080819, 0x19190819,
+    0x2b081908, 0x19190819, 0x2b190808, 0x19190819, 0x08080808, 0x1919082b, 0x08081919, 0x1919082b,
+    0x08082b08, 0x1919082b, 0x08190819, 0x1919082b, 0x08191908, 0x1919082b, 0x082b0808, 0x1919082b,
+    0x19080819, 0x1919082b, 0x19081908, 0x1919082b, 0x19190808, 0x1919082b, 0x192b2b19, 0x1919082b,
+    0x2b080808, 0x1919082b, 0x08080819, 0x19191908, 0x08081908, 0x19191908, 0x0808192b, 0x19191908,
+    0x08082b19, 0x19191908, 0x08190808, 0x19191908, 0x0819082b, 0x19191908, 0x08191919, 0x19191908,
+    0x08192b08, 0x19191908, 0x082b0819, 0x19191908, 0x082b1908, 0x19191908, 0x19080808, 0x19191908,
+    0x1908082b, 0x19191908, 0x19081919, 0x19191908, 0x19082b08, 0x19191908, 0x19190819, 0x19191908,
+    0x19191908, 0x19191908, 0x192b0808, 0x19191908, 0x2b080819, 0x19191908, 0x2b081908, 0x19191908,
+    0x2b190808, 0x19191908, 0x08080808, 0x19191919, 0x0808082b, 0x19191919, 0x08081919, 0x19191919,
+    0x08082b08, 0x19191919, 0x08190819, 0x19191919, 0x08191908, 0x19191919, 0x082b0808, 0x19191919,
+    0x19080819, 0x19191919, 0x19081908, 0x19191919, 0x19190808, 0x19191919, 0x2b080808, 0x19191919,
+    0x08080819, 0x1919192b, 0x08081908, 0x1919192b, 0x08190808, 0x1919192b, 0x082b192b, 0x1919192b,
+    0x19080808, 0x1919192b, 0x08080808, 0x19192b08, 0x0808082b, 0x19192b08, 0x08081919, 0x19192b08,
+    0x08082b08, 0x19192b08, 0x08190819, 0x19192b08, 0x08191908, 0x19192b08, 0x082b0808, 0x19192b08,
+    0x19080819, 0x19192b08, 0x19081908, 0x19192b08, 0x19190808, 0x19192b08, 0x19192b2b, 0x19192b08,
+    0x2b080808, 0x19192b08, 0x08080819, 0x19192b19, 0x08081908, 0x19192b19, 0x08190808, 0x19192b19,
+    0x19080808, 0x19192b19, 0x08080808, 0x19192b2b, 0x08192b19, 0x19192b2b, 0x2b081919, 0x19192b2b,
+    0x2b2b2b08, 0x19192b2b, 0x08080819, 0x192b0808, 0x08081908, 0x192b0808, 0x0808192b, 0x192b0808,
+    0x08190808, 0x192b0808, 0x0819082b, 0x192b0808, 0x08191919, 0x192b0808, 0x08192b08, 0x192b0808,
+    0x082b0819, 0x192b0808, 0x082b1908, 0x192b0808, 0x19080808, 0x192b0808, 0x19081919, 0x192b0808,
+    0x19082b08, 0x192b0808, 0x19190819, 0x192b0808, 0x19191908, 0x192b0808, 0x192b0808, 0x192b0808,
+    0x2b081908, 0x192b0808, 0x2b190808, 0x192b0808, 0x08080808, 0x192b0819, 0x0808082b, 0x192b0819,
+    0x08081919, 0x192b0819, 0x08082b08, 0x192b0819, 0x08190819, 0x192b0819, 0x08191908, 0x192b0819,
+    0x082b0808, 0x192b0819, 0x19080819, 0x192b0819, 0x19081908, 0x192b0819, 0x19190808, 0x192b0819,
+    0x2b080808, 0x192b0819, 0x2b192b19, 0x192b0819, 0x08081908, 0x192b082b, 0x08190808, 0x192b082b,
+    0x19080808, 0x192b082b, 0x1919192b, 0x192b082b, 0x2b2b0819, 0x192b082b, 0x08080808, 0x192b1908,
+    0x08081919, 0x192b1908, 0x08082b08, 0x192b1908, 0x08190819, 0x192b1908, 0x08191908, 0x192b1908,
+    0x082b0808, 0x192b1908, 0x19080819, 0x192b1908, 0x19081908, 0x192b1908, 0x19190808, 0x192b1908,
+    0x2b080808, 0x192b1908, 0x08080819, 0x192b1919, 0x08081908, 0x192b1919, 0x08190808, 0x192b1919,
+    0x19080808, 0x192b1919, 0x19082b2b, 0x192b1919, 0x192b2b08, 0x192b1919, 0x2b19082b, 0x192b1919,
+    0x08080808, 0x192b192b, 0x2b191908, 0x192b192b, 0x08080819, 0x192b2b08, 0x08081908, 0x192b2b08,
+    0x08190808, 0x192b2b08, 0x192b1919, 0x192b2b08, 0x2b192b08, 0x192b2b08, 0x08080808, 0x192b2b19,
+    0x082b2b2b, 0x192b2b19, 0x1908082b, 0x192b2b2b, 0x2b2b0819, 0x192b2b2b, 0x08080808, 0x2b080808,
+    0x0808082b, 0x2b080808, 0x08081919, 0x2b080808, 0x08082b08, 0x2b080808, 0x08190819, 0x2b080808,
+    0x08191908, 0x2b080808, 0x08192b19, 0x2b080808, 0x082b0808, 0x2b080808, 0x082b1919, 0x2b080808,
+    0x19080819, 0x2b080808, 0x19081908, 0x2b080808, 0x19190808, 0x2b080808, 0x1919082b, 0x2b080808,
+    0x19191919, 0x2b080808, 0x19192b08, 0x2b080808, 0x192b0819, 0x2b080808, 0x2b080808, 0x2b080808,
+    0x2b081919, 0x2b080808, 0x2b190819, 0x2b080808, 0x2b191908, 0x2b080808, 0x08080819, 0x2b080819,
+    0x08081908, 0x2b080819, 0x08082b19, 0x2b080819, 0x08190808, 0x2b080819, 0x0819082b, 0x2b080819,
+    0x08191919, 0x2b080819, 0x08192b08, 0x2b080819, 0x082b0819, 0x2b080819, 0x082b1908, 0x2b080819,
+    0x19080808, 0x2b080819, 0x1908082b, 0x2b080819, 0x19081919, 0x2b080819, 0x19082b08, 0x2b080819,
+    0x19190819, 0x2b080819, 0x19191908, 0x2b080819, 0x2b080819, 0x2b080819, 0x2b081908, 0x2b080819,
+    0x2b190808, 0x2b080819, 0x2b2b2b19, 0x2b080819, 0x08080808, 0x2b08082b, 0x08081919, 0x2b08082b,
+    0x08082b2b, 0x2b08082b, 0x08190819, 0x2b08082b, 0x08191908, 0x2b08082b, 0x19080819, 0x2b08082b,
+    0x19081908, 0x2b08082b, 0x19190808, 0x2b08082b, 0x08080819, 0x2b081908, 0x08081908, 0x2b081908,
+    0x0808192b, 0x2b081908, 0x08082b19, 0x2b081908, 0x08190808, 0x2b081908, 0x0819082b, 0x2b081908,
+    0x08191919, 0x2b081908, 0x08192b08, 0x2b081908, 0x082b0819, 0x2b081908, 0x19080808, 0x2b081908,
+    0x1908082b, 0x2b081908, 0x19081919, 0x2b081908, 0x19082b08, 0x2b081908, 0x19190819, 0x2b081908,
+    0x19191908, 0x2b081908, 0x192b0808, 0x2b081908, 0x2b080819, 0x2b081908, 0x2b081908, 0x2b081908,
+    0x2b190808, 0x2b081908, 0x08080808, 0x2b081919, 0x0808082b, 0x2b081919, 0x08081919, 0x2b081919,
+    0x08082b08, 0x2b081919, 0x08190819, 0x2b081919, 0x08191908, 0x2b081919, 0x082b0808, 0x2b081919,
+    0x19080819, 0x2b081919, 0x19081908, 0x2b081919, 0x19190808, 0x2b081919, 0x2b080808, 0x2b081919,
+    0x2b082b2b, 0x2b081919, 0x08080819, 0x2b08192b, 0x08081908, 0x2b08192b, 0x08190808, 0x2b08192b,
+    0x082b2b19, 0x2b08192b, 0x19080808, 0x2b08192b, 0x08080808, 0x2b082b08, 0x08081919, 0x2b082b08,
+    0x08190819, 0x2b082b08, 0x08191908, 0x2b082b08, 0x19080819, 0x2b082b08, 0x19081908, 0x2b082b08,
+    0x19190808, 0x2b082b08, 0x2b2b082b, 0x2b082b08, 0x08080819, 0x2b082b19, 0x08081908, 0x2b082b19,
+    0x19080808, 0x2b082b19, 0x192b1919, 0x2b082b19, 0x082b082b, 0x2b082b2b, 0x19192b08, 0x2b082b2b,
+    0x19192b2b, 0x2b082b2b, 0x2b08082b, 0x2b082b2b, 0x2b2b082b, 0x2b082b2b, 0x08080819, 0x2b190808,
+    0x08081908, 0x2b190808, 0x08082b19, 0x2b190808, 0x08190808, 0x2b190808, 0x0819082b, 0x2b190808,
+    0x08191919, 0x2b190808, 0x08192b08, 0x2b190808, 0x082b1908, 0x2b190808, 0x19080808, 0x2b190808,
+    0x1908082b, 0x2b190808, 0x19081919, 0x2b190808, 0x19082b08, 0x2b190808, 0x19190819, 0x2b190808,
+    0x19191908, 0x2b190808, 0x192b0808, 0x2b190808, 0x2b080819, 0x2b190808, 0x2b081908, 0x2b190808,
+    0x2b190808, 0x2b190808, 0x08080808, 0x2b190819, 0x08081919, 0x2b190819, 0x08190819, 0x2b190819,
+    0x08191908, 0x2b190819, 0x19080819, 0x2b190819, 0x19081908, 0x2b190819, 0x19190808, 0x2b190819,
+    0x19192b2b, 0x2b190819, 0x08080819, 0x2b19082b, 0x08081908, 0x2b19082b, 0x08190808, 0x2b19082b,
+    0x19080808, 0x2b19082b, 0x2b2b192b, 0x2b19082b, 0x08080808, 0x2b191908, 0x0808082b, 0x2b191908,
+    0x08081919, 0x2b191908, 0x08082b08, 0x2b191908, 0x08190819, 0x2b191908, 0x08191908, 0x2b191908,
+    0x082b0808, 0x2b191908, 0x19080819, 0x2b191908, 0x19081908, 0x2b191908, 0x19190808, 0x2b191908,
+    0x2b080808, 0x2b191908, 0x2b19192b, 0x2b191908, 0x08080819, 0x2b191919, 0x08081908, 0x2b191919,
+    0x08190808, 0x2b191919, 0x19080808, 0x2b191919, 0x2b192b08, 0x2b191919, 0x2b2b0819, 0x2b191919,
+    0x08080808, 0x2b19192b, 0x1908192b, 0x2b19192b, 0x192b1908, 0x2b19192b, 0x08080819, 0x2b192b08,
+    0x08081908, 0x2b192b08, 0x08190808, 0x2b192b08, 0x082b192b, 0x2b192b08, 0x19080808, 0x2b192b08,
+    0x2b2b2b19, 0x2b192b08, 0x08080808, 0x2b192b19, 0x19082b19, 0x2b192b19, 0x1919082b, 0x2b192b19,
+    0x2b190808, 0x2b192b2b, 0x08080808, 0x2b2b0808, 0x08081919, 0x2b2b0808, 0x08082b2b, 0x2b2b0808,
+    0x08191908, 0x2b2b0808, 0x082b082b, 0x2b2b0808, 0x082b2b2b, 0x2b2b0808, 0x19080819, 0x2b2b0808,
+    0x19081908, 0x2b2b0808, 0x19190808, 0x2b2b0808, 0x2b2b082b, 0x2b2b0808, 0x2b2b2b2b, 0x2b2b0808,
+    0x19080808, 0x2b2b0819, 0x192b1919, 0x2b2b0819, 0x0808082b, 0x2b2b082b, 0x08082b2b, 0x2b2b082b,
+    0x082b082b, 0x2b2b082b, 0x082b2b08, 0x2b2b082b, 0x082b2b2b, 0x2b2b082b, 0x2b08082b, 0x2b2b082b,
+    0x2b082b08, 0x2b2b082b, 0x2b082b2b, 0x2b2b082b, 0x2b2b2b08, 0x2b2b082b, 0x08080819, 0x2b2b1908,
+    0x08081908, 0x2b2b1908, 0x08190808, 0x2b2b1908, 0x19080808, 0x2b2b1908, 0x2b082b19, 0x2b2b1908,
+    0x2b2b1908, 0x2b2b1908, 0x08080808, 0x2b2b1919, 0x08192b19, 0x2b2b1919, 0x19190819, 0x2b2b192b,
+    0x08082b2b, 0x2b2b2b08, 0x082b2b08, 0x2b2b2b08, 0x2b2b082b, 0x2b2b2b08, 0x19191908, 0x2b2b2b19,
+    0x2b08192b, 0x2b2b2b19, 0x08082b08, 0x2b2b2b2b, 0x08082b2b, 0x2b2b2b2b, 0x082b0808, 0x2b2b2b2b,
+    0x082b082b, 0x2b2b2b2b, 0x082b2b08, 0x2b2b2b2b, 0x2b082b08, 0x2b2b2b2b, 0x2b2b2b2b, 0x2b2b2b2b
+);
+
+struct iq2_s {
+    d: f16,
+    qs: array<f16, 32>,
+    qh: array<f16, 4>,
+    scales: array<f16, 4>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    var src1_i = src1_idx_base + offset * 256;
+    var qs_vals : array<u32, 16>;
+    for (var i: u32 = 0; i < 16; i++) {
+        qs_vals[i] = bitcast<u32>(vec2(block.qs[i * 2], block.qs[i * 2 + 1]));
+    }
+    var qh_vals = array<u32, 2>(
+        bitcast<u32>(vec2(block.qh[0], block.qh[1])),
+        bitcast<u32>(vec2(block.qh[2], block.qh[3]))
+    );
+    var scale_vals = array<u32, 2>(
+        bitcast<u32>(vec2(block.scales[0], block.scales[1])),
+        bitcast<u32>(vec2(block.scales[2], block.scales[3]))
+    );
+    var sum = 0.0;
+    for (var ib: u32 = 0; ib < 8; ib ++) {
+        let s = get_byte(scale_vals[ib / 4], ib % 4);
+        let db = array<f32, 2>(
+            d * (0.5 + f32(s & 0xF)) * 0.25,
+            d * (0.5 + f32(s >> 4)) * 0.25
+        );
+        let qs_w = qs_vals[ib];
+        for (var l: u32 = 0; l < 4; l++) {
+            let qh_b = (get_byte(qh_vals[ib / 4], ib % 4) << (8 - 2 * l)) & 0x300;
+            let ig = (get_byte(qs_w, l) | qh_b) * 8;
+            let signs = get_byte(qs_vals[ib + 8], l);
+            let dl = db[l/2];
+            for (var j: u32 = 0; j < 8; j++) {
+                let g = get_byte(iq2s_grid[(ig + j) / 4], (ig + j) % 4);
+                let m = select(1.0, -1.0, (get_byte(kmask_iq2xs[j / 4], j % 4) & signs) != 0);
+                sum += dl * f32(g) * m * src1[src1_i];
+                src1_i++;
+            }
+        }
+    }
+    return sum;
+}
+
+
+#enddecl(IQ2_S)
+
+#decl(IQ3_XSS)
+
+const iq3xxs_grid = array<u32, 256>(
+    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
+    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
+    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
+    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
+    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
+    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
+    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
+    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
+    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
+    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
+    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
+    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
+    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
+    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
+    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
+    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
+    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
+    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
+    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
+    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
+    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
+    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
+    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
+    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
+    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
+    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
+    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
+    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
+    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
+    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
+    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
+    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04
+);
+
+struct iq3_xxs {
+    d: f16,
+    qs: array<f16, 48>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    var src1_i = src1_idx_base + offset * 256;
+    var sum = 0.0;
+    for (var ib: u32 = 0; ib < 16; ib += 2) {
+        let sc_sign = bitcast<u32>(vec2(block.qs[ib + 32], block.qs[ib + 33]));
+        let db = d * (0.5 + f32(sc_sign >> 28)) * 0.5;
+        for (var l: u32 = 0; l < 4; l++) {
+            let is = (sc_sign >> (7 * l)) & 127;
+            let signs = get_byte(ksigns_iq2xs[is / 4], is % 4);
+            let ig_val = bitcast<u32>(vec2(block.qs[ib * 2 + l], 0.0));
+            let ig1 = get_byte(ig_val, 0);
+            let ig2 = get_byte(ig_val, 1);
+            for (var j: u32 = 0; j < 4; j++) {
+                let g1 = get_byte(iq3xxs_grid[ig1], j);
+                let g2 = get_byte(iq3xxs_grid[ig2], j);
+                let m1 = select(1.0, -1.0, (get_byte(kmask_iq2xs[0], j) & signs) != 0);
+                let m2 = select(1.0, -1.0, (get_byte(kmask_iq2xs[1], j) & signs) != 0);
+                sum += db * f32(g1) * m1 * src1[src1_i];
+                sum += db * f32(g2) * m2 * src1[src1_i + 4];
+                src1_i++;
+            }
+            src1_i += 4;
+        }
+    }
+    return sum;
+}
+
+#enddecl(IQ3_XSS)
+
+#decl(IQ3_S)
+
+const iq3s_grid = array<u32, 512>(
+    0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
+    0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
+    0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
+    0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
+    0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
+    0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
+    0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
+    0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
+    0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
+    0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
+    0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
+    0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
+    0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
+    0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
+    0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
+    0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
+    0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
+    0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
+    0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
+    0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
+    0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
+    0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
+    0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
+    0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
+    0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
+    0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
+    0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
+    0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
+    0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
+    0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
+    0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
+    0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
+    0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
+    0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
+    0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
+    0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
+    0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
+    0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
+    0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
+    0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
+    0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
+    0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
+    0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
+    0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
+    0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
+    0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
+    0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
+    0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
+    0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
+    0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
+    0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
+    0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
+    0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
+    0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
+    0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
+    0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
+    0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
+    0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
+    0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
+    0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
+    0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
+    0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
+    0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
+    0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101
+);
+
+struct iq3_s {
+    d: f16,
+    qs: array<f16, 32>,
+    qh: array<f16, 4>,
+    signs: array<f16, 16>,
+    scales: array<f16, 2>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    var src1_i = src1_idx_base + offset * 256;
+    var qh_vals = array<u32, 2>(
+        bitcast<u32>(vec2(block.qh[0], block.qh[1])),
+        bitcast<u32>(vec2(block.qh[2], block.qh[3]))
+    );
+    var sign_vals: array<u32, 8>;
+    for (var i: u32 = 0; i < 8; i++) {
+        sign_vals[i] = bitcast<u32>(vec2(block.signs[i * 2], block.signs[i * 2 + 1]));
+    }
+    var scale_vals = bitcast<u32>(vec2(block.scales[0], block.scales[1]));
+    var sum = 0.0;
+    for (var ib: u32 = 0; ib < 4; ib++) {
+        let s = get_byte(scale_vals, ib);
+        let db = array<f32, 2>(
+            d * (1.0 + 2.0 * f32(s & 0xF)),
+            d * (1.0 + 2.0 * f32(s >> 4))
+        );
+        for (var k: u32 = 0; k < 2; k++) {
+            let dl = db[k];
+            let qh_byte = get_byte(qh_vals[ib / 2], (ib % 2) * 2 + k);
+            let sign_w = sign_vals[ib * 2 + k];
+            for (var l: u32 = 0; l < 4; l++) {
+                let signs = get_byte(sign_w, l);
+                let ig_val = bitcast<u32>(vec2(block.qs[ib * 8 + k * 4 + l], 0.0));
+                let ig1 = get_byte(ig_val, 0) | ((qh_byte << ((8 - (2 * l)))) & 256);
+                let ig2 = get_byte(ig_val, 1) | ((qh_byte << ((7 - (2 * l)))) & 256);
+                for (var j: u32 = 0; j < 4; j++) {
+                    let g1 = get_byte(iq3s_grid[ig1], j);
+                    let g2 = get_byte(iq3s_grid[ig2], j);
+                    let m1 = select(1.0, -1.0, (get_byte(kmask_iq2xs[0], j) & signs) != 0);
+                    let m2 = select(1.0, -1.0, (get_byte(kmask_iq2xs[1], j) & signs) != 0);
+                    sum += dl * f32(g1) * m1 * src1[src1_i];
+                    sum += dl * f32(g2) * m2 * src1[src1_i + 4];
+                    src1_i++;
+                }
+                src1_i += 4;
+            }
+        }
+    }
+    return sum;
+}
+#enddecl(IQ3_S)
+
+#decl(IQ1_TABLE)
+
+const IQ1_DELTA: f32 = 0.125;
+
+const iq1_grid = array<u32, 1024>(
+    0xfffdffff, 0xfff7fff0, 0xffccfff5, 0xffdfffc0, 0xffd7ffdd, 0xff30ffd5, 0xff03ff0c, 0xff10ff01,
+    0xff7dff7f, 0xff75ff77, 0xff5fff40, 0xff57ff5d, 0xfcf3ff55, 0xfcccfcf0, 0xfcc1fcc3, 0xfcc5fcc4,
+    0xfc3cfcd0, 0xfc34fc31, 0xfc00fc0d, 0xfc1cfc05, 0xfc11fc13, 0xfc70fc17, 0xfc43fc4c, 0xfc50fc41,
+    0xfdfdfdff, 0xfdf5fdf7, 0xfddffdc0, 0xfdd7fddd, 0xfd30fdd5, 0xfd04fd0c, 0xfd14fd13, 0xfd7dfd7f,
+    0xfd75fd77, 0xfd40fd4c, 0xfd5ffd44, 0xfd57fd5d, 0xf3ccfd55, 0xf3c1f3c3, 0xf33cf3d0, 0xf300f334,
+    0xf313f305, 0xf34cf310, 0xf350f344, 0xf0f3f0fc, 0xf0f1f0f0, 0xf0c7f0c0, 0xf0d4f0c5, 0xf030f03f,
+    0xf00ff035, 0xf003f00c, 0xf001f000, 0xf01ff004, 0xf010f01d, 0xf015f017, 0xf04cf07c, 0xf047f040,
+    0xf05cf045, 0xf050f053, 0xf054f051, 0xf1c4f1c3, 0xf133f13c, 0xf10df10f, 0xf107f100, 0xf11cf11f,
+    0xf114f111, 0xf14cf170, 0xf144f143, 0xf7fdf7ff, 0xf7f5f7f7, 0xf7dff7c0, 0xf7d7f7dd, 0xf730f7d5,
+    0xf701f70c, 0xf77ff710, 0xf777f77d, 0xf740f775, 0xf75df75f, 0xf755f757, 0xf4ccf4f0, 0xf4c4f4c3,
+    0xf4d0f4d3, 0xf40ff43c, 0xf400f40c, 0xf413f41c, 0xf44cf414, 0xf441f443, 0xf450f444, 0xf5fdf5ff,
+    0xf5f5f5f7, 0xf5dff5c0, 0xf5d7f5dd, 0xf530f5d5, 0xf504f50c, 0xf510f51c, 0xf57df57f, 0xf577f570,
+    0xf540f575, 0xf55df55f, 0xf555f557, 0xcfcccfcf, 0xcfc4cfc3, 0xcfd0cfd3, 0xcf33cf3c, 0xcf00cf0f,
+    0xcf1ccf07, 0xcf10cf13, 0xcf4ccf14, 0xcf41cf43, 0xcf50cf5c, 0xccf3ccfc, 0xccf4ccf1, 0xcccdcccf,
+    0xccc7ccc0, 0xccd3ccdc, 0xcc30ccd4, 0xcc0fcc35, 0xcc0dcc0c, 0xcc00cc03, 0xcc04cc01, 0xcc10cc1f,
+    0xcc4dcc73, 0xcc5ccc40, 0xcdcccc53, 0xcdc1cdc3, 0xcd3fcdd0, 0xcd34cd31, 0xcd00cd0d, 0xcd05cd07,
+    0xcd11cd13, 0xcd4ccd70, 0xcd41cd43, 0xc3fccd50, 0xc3f4c3f1, 0xc3c0c3c3, 0xc3c4c3c7, 0xc3d1c3dc,
+    0xc330c33c, 0xc337c331, 0xc30cc335, 0xc300c303, 0xc304c301, 0xc310c31d, 0xc373c317, 0xc34fc374,
+    0xc340c343, 0xc344c347, 0xc35cc345, 0xc350c353, 0xc0fdc354, 0xc0f5c0f0, 0xc0c3c0cc, 0xc0c1c0c0,
+    0xc0dfc0c4, 0xc0d0c0dd, 0xc0d5c0d7, 0xc033c03c, 0xc031c030, 0xc00dc00c, 0xc000c003, 0xc004c001,
+    0xc01cc005, 0xc010c013, 0xc014c011, 0xc07dc07f, 0xc070c073, 0xc075c077, 0xc04cc04f, 0xc040c043,
+    0xc044c041, 0xc05fc045, 0xc050c05d, 0xc1f3c1fc, 0xc1f1c1f0, 0xc1c1c1c0, 0xc1c5c1c7, 0xc1d1c1dc,
+    0xc13dc13f, 0xc130c133, 0xc135c137, 0xc100c10c, 0xc107c101, 0xc11cc104, 0xc110c113, 0xc114c117,
+    0xc171c115, 0xc14dc175, 0xc153c140, 0xc7ccc154, 0xc7d0c7c1, 0xc733c73c, 0xc734c731, 0xc700c70f,
+    0xc705c707, 0xc71cc71f, 0xc711c713, 0xc770c714, 0xc743c74c, 0xc4cfc750, 0xc4c0c4cd, 0xc4dcc4c5,
+    0xc43dc4d0, 0xc430c433, 0xc40cc437, 0xc400c403, 0xc404c401, 0xc41fc405, 0xc415c410, 0xc44cc474,
+    0xc440c44d, 0xc45cc447, 0xc454c451, 0xc5c1c5f4, 0xc5d1c5d3, 0xc531c533, 0xc50fc534, 0xc500c50d,
+    0xc51cc507, 0xc514c511, 0xc54cc570, 0xc545c541, 0xdffddfff, 0xdff5dff7, 0xdfdfdfc0, 0xdfd0dfdd,
+    0xdfd5dfd7, 0xdf0cdf30, 0xdf1cdf04, 0xdf7fdf10, 0xdf77df7d, 0xdf40df75, 0xdf5ddf5f, 0xdf57df50,
+    0xdcf0df55, 0xdcc3dccc, 0xdcd0dcc4, 0xdc33dc3d, 0xdc00dc34, 0xdc05dc07, 0xdc13dc1c, 0xdc11dc10,
+    0xdc4fdc70, 0xdc44dc41, 0xddfcdc50, 0xddf5ddf7, 0xddc0ddcc, 0xdddddddf, 0xddd5ddd7, 0xdd0cdd30,
+    0xdd04dd01, 0xdd7cdd10, 0xdd75dd77, 0xdd40dd4c, 0xdd5ddd5f, 0xdd55dd57, 0xd3c3d3f0, 0xd3c4d3c1,
+    0xd333d3d0, 0xd331d330, 0xd30dd334, 0xd307d300, 0xd311d305, 0xd34cd370, 0xd344d343, 0xd350d35c,
+    0xd0c0d0f4, 0xd0d4d0dc, 0xd030d03f, 0xd00cd037, 0xd000d003, 0xd01dd004, 0xd017d010, 0xd04fd074,
+    0xd040d043, 0xd045d047, 0xd053d05c, 0xd054d051, 0xd1cfd1f0, 0xd1c4d1cd, 0xd13cd1d0, 0xd100d134,
+    0xd11cd11f, 0xd173d114, 0xd14fd171, 0xd7ffd145, 0xd7f7d7fd, 0xd7c0d7f5, 0xd7ddd7df, 0xd7d5d7d7,
+    0xd70cd730, 0xd710d703, 0xd77dd77f, 0xd775d777, 0xd75dd75f, 0xd755d757, 0xd4ccd4f4, 0xd4c4d4c3,
+    0xd431d4d0, 0xd40dd434, 0xd41cd400, 0xd411d413, 0xd470d414, 0xd441d44f, 0xd453d444, 0xd5ffd450,
+    0xd5f7d5fd, 0xd5dfd5f5, 0xd5d7d5dd, 0xd530d5d5, 0xd501d50c, 0xd510d504, 0xd57dd57f, 0xd575d577,
+    0xd55fd540, 0xd557d55d, 0x3ff0d555, 0x3fc13fcc, 0x3f343fd0, 0x3f003f0d, 0x3f053f07, 0x3f133f1c,
+    0x3f433f11, 0x3f5c3f44, 0x3cff3f51, 0x3cf33cfc, 0x3cf43cf1, 0x3cc03ccd, 0x3cc73cc1, 0x3cdc3cc5,
+    0x3cd43cd1, 0x3c373c30, 0x3c0c3c35, 0x3c003c03, 0x3c043c01, 0x3c103c05, 0x3c153c17, 0x3c733c7c,
+    0x3c4f3c71, 0x3c403c4d, 0x3c5c3c5f, 0x3df03c5d, 0x3dc33dcc, 0x3dd03dc1, 0x3d0d3d3c, 0x3d053d00,
+    0x3d143d13, 0x3d433d74, 0x33fc3d50, 0x33c433c0, 0x333033d4, 0x33353337, 0x3303330c, 0x33013300,
+    0x331d331c, 0x33173310, 0x337c3315, 0x33743371, 0x334d334f, 0x335f3340, 0x3354335c, 0x30fd30fc,
+    0x30f530f0, 0x30c330cc, 0x30c130c0, 0x30df30c4, 0x30d530d0, 0x3033303c, 0x30313030, 0x300f3034,
+    0x3003300c, 0x30013000, 0x30043007, 0x3013301c, 0x30113010, 0x307d3014, 0x30703073, 0x304c3077,
+    0x30403043, 0x30443041, 0x30503045, 0x30553057, 0x31f031fc, 0x31c331f4, 0x31c731c0, 0x31dc31c5,
+    0x31d431d3, 0x313d313f, 0x31373130, 0x310c310f, 0x3100310d, 0x31043101, 0x3110311d, 0x317c3117,
+    0x31753170, 0x31403143, 0x3153315c, 0x37f03151, 0x37c037cc, 0x37d037c5, 0x3734373d, 0x3700370f,
+    0x371c3707, 0x37113713, 0x37703714, 0x3743374c, 0x37443741, 0x34fc3750, 0x34f134f0, 0x34cf34f5,
+    0x34c034c3, 0x34dc34c7, 0x34d134d3, 0x3430343f, 0x340c3435, 0x3403340d, 0x34013400, 0x341f3404,
+    0x3410341d, 0x34153411, 0x34743471, 0x3440344d, 0x34473441, 0x3453345c, 0x34543451, 0x353335c1,
+    0x35343531, 0x35073500, 0x35133505, 0x35433514, 0x0ffc3550, 0x0ff00ff3, 0x0ff40ff1, 0x0fc00fcd,
+    0x0fdc0fc5, 0x0fd40fd3, 0x0f300f3f, 0x0f0c0f37, 0x0f000f03, 0x0f040f01, 0x0f170f10, 0x0f740f71,
+    0x0f470f40, 0x0f5c0f5f, 0x0f540f51, 0x0cf70cf0, 0x0cf50cf4, 0x0cc30ccc, 0x0cc10cc0, 0x0cc40cc7,
+    0x0cd00cdf, 0x0cd70cd1, 0x0c3c0cd5, 0x0c300c33, 0x0c340c31, 0x0c0c0c0f, 0x0c030c0d, 0x0c010c00,
+    0x0c040c07, 0x0c1c0c05, 0x0c100c13, 0x0c140c11, 0x0c700c7d, 0x0c430c4c, 0x0c410c40, 0x0c5f0c44,
+    0x0c550c50, 0x0df10dfc, 0x0dc00dcd, 0x0ddc0dc5, 0x0d3d0dd3, 0x0d350d30, 0x0d030d0c, 0x0d010d00,
+    0x0d1d0d04, 0x0d700d10, 0x0d4d0d4f, 0x0d440d40, 0x0d530d45, 0x03f003f3, 0x03c303cc, 0x03c103c0,
+    0x03c403c7, 0x03d003dc, 0x03d503d7, 0x0333033c, 0x03310330, 0x03350334, 0x030c030f, 0x03000303,
+    0x03070301, 0x03050304, 0x031d031c, 0x03100313, 0x03140311, 0x0377037f, 0x034c0375, 0x03400343,
+    0x03440341, 0x0353035c, 0x03550350, 0x00fd00fc, 0x00f000f3, 0x00f400f1, 0x00cc00cf, 0x00c300cd,
+    0x00c100c0, 0x00c500c4, 0x00d300dc, 0x00d100d0, 0x003f00d4, 0x003d003c, 0x00300033, 0x00370031,
+    0x000f0034, 0x000d000c, 0x00000003, 0x00070001, 0x00050004, 0x001c001f, 0x00100013, 0x00170011,
+    0x00150014, 0x0073007c, 0x00740070, 0x004f0075, 0x0043004c, 0x00410040, 0x00440047, 0x0053005c,
+    0x00510050, 0x01ff0054, 0x01fd01fc, 0x01f101f3, 0x01f401f7, 0x01c301cc, 0x01c701c0, 0x01df01c4,
+    0x01dd01dc, 0x01d001d3, 0x01d701d1, 0x013c01d4, 0x01310130, 0x01340137, 0x010f0135, 0x010d010c,
+    0x01000103, 0x01070101, 0x01050104, 0x0113011c, 0x01140110, 0x0170017d, 0x01770171, 0x01750174,
+    0x0140014c, 0x015d0145, 0x01510150, 0x01540157, 0x07f007f3, 0x07f407f1, 0x07c007cf, 0x07dc07c7,
+    0x073007d5, 0x07350737, 0x0703070c, 0x07010700, 0x07040707, 0x071d071f, 0x07100713, 0x0774077d,
+    0x074d074f, 0x07470740, 0x0754075c, 0x04fd04fc, 0x04f504f0, 0x04c304cc, 0x04c104c0, 0x04d004c4,
+    0x0433043c, 0x04310430, 0x040f0434, 0x040d040c, 0x04000403, 0x04070401, 0x04050404, 0x0413041c,
+    0x04110410, 0x047c0414, 0x04740470, 0x0443044c, 0x04410440, 0x04440447, 0x05f30450, 0x05c005f7,
+    0x05df05c5, 0x05d105d0, 0x053005d4, 0x05340537, 0x0500050c, 0x05070501, 0x051d0504, 0x05170510,
+    0x057c0515, 0x054d0575, 0x05410540, 0x05450547, 0x1ff0055c, 0x1fc11fc3, 0x1fd01fc4, 0x1f0f1f33,
+    0x1f011f00, 0x1f051f07, 0x1f131f1c, 0x1f141f11, 0x1f411f7c, 0x1cfc1f50, 0x1cf11cf3, 0x1ccd1cf4,
+    0x1cdc1cc0, 0x1cd11cdd, 0x1c301cd4, 0x1c0c1c34, 0x1c011c00, 0x1c101c04, 0x1c151c11, 0x1c751c73,
+    0x1c401c4d, 0x1c511c5c, 0x1dcc1c54, 0x1dc41dc1, 0x1d3c1d3f, 0x1d001d31, 0x1d071d01, 0x1d701d1f,
+    0x1d411d4c, 0x13cc1d50, 0x13c013cd, 0x13c513c1, 0x13d113dc, 0x133f13d4, 0x1330133d, 0x13351337,
+    0x1303130c, 0x13011300, 0x13051304, 0x131d131f, 0x13731310, 0x13741370, 0x134d134f, 0x13401343,
+    0x13471341, 0x135c1345, 0x13541353, 0x10f710f0, 0x10cc10f5, 0x10c110c0, 0x103310c4, 0x10311030,
+    0x100f1034, 0x1003100c, 0x10011000, 0x101c1004, 0x10101013, 0x10141011, 0x10741071, 0x104c1075,
+    0x10411040, 0x10451044, 0x1050105d, 0x10571051, 0x11f411fd, 0x11df11c0, 0x11d711d1, 0x113f11d4,
+    0x11371130, 0x110c1135, 0x11001103, 0x11071101, 0x111f1105, 0x11171110, 0x117d117f, 0x11751170,
+    0x11411143, 0x11441147, 0x1153115f, 0x11551151, 0x17c417c1, 0x173c17d0, 0x1700170d, 0x171c1705,
+    0x17701714, 0x1747174c, 0x14fc1751, 0x14cf14f3, 0x14dc14c0, 0x14d114d3, 0x143f14d4, 0x1430143c,
+    0x14371431, 0x1403140c, 0x14011400, 0x141f1404, 0x14151410, 0x1473147d, 0x14401475, 0x1453145c,
+    0x14541450, 0x15c115cc, 0x153c15c7, 0x15341533, 0x1500150f, 0x15051507, 0x15101513, 0x15711514,
+    0x15471543, 0x15511545, 0x7ffd7fff, 0x7ff57ff7, 0x7fdd7fdf, 0x7fd57fd7, 0x7f0f7f30, 0x7f037f0c,
+    0x7f047f01, 0x7f7f7f10, 0x7f777f7d, 0x7f407f75, 0x7f5d7f5f, 0x7f557f57, 0x7ccc7cf0, 0x7cc17cc3,
+    0x7cd07cc4, 0x7c337c3c, 0x7c0f7c34, 0x7c007c0d, 0x7c077c01, 0x7c137c04, 0x7c147c11, 0x7c747c70,
+    0x7c417c43, 0x7c507c44, 0x7dfd7dff, 0x7df57df7, 0x7ddf7dc0, 0x7dd77ddd, 0x7d0c7dd5, 0x7d047d03,
+    0x7d7f7d10, 0x7d777d7d, 0x7d407d75, 0x7d5d7d5f, 0x7d557d57, 0x73c473c3, 0x7333733c, 0x7300730c,
+    0x731c7305, 0x73147313, 0x73447343, 0x70f470fc, 0x70c070cd, 0x70d170c5, 0x703f70d4, 0x7030703c,
+    0x700c7037, 0x70007003, 0x70047001, 0x70107005, 0x70177011, 0x707c7015, 0x70717073, 0x704f7074,
+    0x7040704d, 0x70517047, 0x71c171cc, 0x71d071c4, 0x7133713c, 0x71357134, 0x7100710f, 0x71057104,
+    0x7111711c, 0x71707115, 0x7145714c, 0x77ff7153, 0x77f777fd, 0x77c077f5, 0x77dd77df, 0x77d577d7,
+    0x7730773c, 0x7703770c, 0x77107704, 0x777f7714, 0x7777777d, 0x77407775, 0x775d775f, 0x77557757,
+    0x74f174f0, 0x74c374cc, 0x74d074c1, 0x7433743c, 0x74347431, 0x740d740f, 0x74057400, 0x7413741c,
+    0x74417470, 0x74507444, 0x75fd75ff, 0x75f575f7, 0x75df75c0, 0x75d775dd, 0x753075d5, 0x7503750c,
+    0x757f7501, 0x7577757d, 0x75407575, 0x755d755f, 0x75557557, 0x4fcc4ff0, 0x4fc74fc1, 0x4fd04fc4,
+    0x4f314f3c, 0x4f004f34, 0x4f054f07, 0x4f154f14, 0x4f4c4f70, 0x4f414f43, 0x4f504f44, 0x4cf34cfc,
+    0x4cf44cf1, 0x4cc04ccf, 0x4cc54cc7, 0x4cd34cdc, 0x4cd44cd1, 0x4c304c3f, 0x4c0c4c0f, 0x4c004c03,
+    0x4c044c01, 0x4c104c1d, 0x4c714c73, 0x4c404c4d, 0x4c5c4c47, 0x4c514c53, 0x4df04c54, 0x4dc34dcc,
+    0x4dd04dc4, 0x4d314d33, 0x4d0f4d34, 0x4d004d0d, 0x4d114d07, 0x4d704d14, 0x4d414d43, 0x43fc4d54,
+    0x43f143f3, 0x43c043cf, 0x43d143c7, 0x4335433f, 0x4303430c, 0x43014300, 0x43044307, 0x431c431f,
+    0x4310431d, 0x43714373, 0x4343434d, 0x43474340, 0x4354435c, 0x40f040ff, 0x40f540f7, 0x40cc40cf,
+    0x40c040c3, 0x40c440c1, 0x40d040dc, 0x40d540d4, 0x4033403c, 0x40314030, 0x400f4034, 0x400d400c,
+    0x40004003, 0x40074001, 0x40054004, 0x4013401c, 0x40114010, 0x407c4014, 0x40774070, 0x404d404c,
+    0x40404043, 0x40444041, 0x405f4045, 0x4050405d, 0x40554057, 0x41f341fc, 0x41c041cf, 0x41df41c4,
+    0x41d441d1, 0x41374130, 0x410c4134, 0x4100410d, 0x41044101, 0x41174110, 0x4173417d, 0x41754174,
+    0x4143414d, 0x41534140, 0x41544151, 0x47c147f0, 0x47d047c4, 0x4731473c, 0x470d470f, 0x47014700,
+    0x47134705, 0x47704710, 0x4741474c, 0x47504744, 0x44f144f3, 0x44cf44f4, 0x44c044cd, 0x44c544c7,
+    0x44dc44df, 0x44d144d3, 0x443d443f, 0x44374430, 0x440c4435, 0x44004403, 0x44044401, 0x4410441d,
+    0x44154411, 0x4473447c, 0x444d444f, 0x44454440, 0x4451445c, 0x45c045f0, 0x453345d0, 0x45344531,
+    0x4500450f, 0x451c4507, 0x454c4570, 0x45404543, 0x5fff4541, 0x5ff75ffd, 0x5fc05ff5, 0x5fdd5fdf,
+    0x5fd55fd7, 0x5f0c5f30, 0x5f015f03, 0x5f7f5f04, 0x5f775f7d, 0x5f405f75, 0x5f5d5f5f, 0x5f555f57,
+    0x5cf45cf0, 0x5cc35ccc, 0x5cc45cc1, 0x5c315cc5, 0x5c0c5c34, 0x5c075c00, 0x5c1c5c05, 0x5c705c13,
+    0x5c4d5c4f, 0x5c445c41, 0x5df75dfd, 0x5dcf5df5, 0x5ddd5dc4, 0x5dd55dd7, 0x5d0c5d30, 0x5d045d01,
+    0x5d7f5d10, 0x5d775d7d, 0x5d405d75, 0x5d5d5d5f, 0x5d555d57, 0x53d053c4, 0x5333533c, 0x5303530f,
+    0x53075300, 0x531c5305, 0x53115310, 0x53145317, 0x50f15370, 0x50cf50f4, 0x50c050cd, 0x50d150c7,
+    0x503d50d4, 0x500c5030, 0x50005003, 0x50045001, 0x50155010, 0x5073507c, 0x50715070, 0x504d5074,
+    0x50475040, 0x51cc51f0, 0x51c551c1, 0x51d051dc, 0x51315133, 0x510d5135, 0x51015100, 0x511f5107,
+    0x5171511d, 0x5140514f, 0x51445141, 0x5153515c, 0x57ff5151, 0x57f757fd, 0x57df57f5, 0x57d757dd,
+    0x570c57d5, 0x57015703, 0x577f5704, 0x5777577d, 0x57405775, 0x575d575f, 0x57555757, 0x54c354f0,
+    0x54dc54c4, 0x543c54d0, 0x5400540f, 0x541c5405, 0x54145411, 0x5441544f, 0x55fd55ff, 0x55f555f7,
+    0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557
+);
+
+#enddecl(IQ1_TABLE)
+
+#decl(IQ1_S)
+
+struct iq1_s {
+    d: f16,
+    qs: array<f16, 16>,
+    qh: array<f16, 8>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    var src1_i = src1_idx_base + offset * 256;
+    var sum = 0.0;
+    for (var ib: u32 = 0; ib < 8; ib++) {
+        let qh = bitcast<u32>(vec2(block.qh[ib], 0.0));
+        let dl = d * (2 * f32((qh >> 12) & 7) + 1);
+        let delta = select(IQ1_DELTA, -IQ1_DELTA, (qh & 0x8000) != 0);
+        let qs_w = bitcast<u32>(vec2(block.qs[ib * 2], block.qs[ib * 2 + 1]));
+        for (var l: u32 = 0; l < 4; l++) {
+            let ig = (get_byte(qs_w, l) | (((qh >> (3 * l)) & 7) << 8)) * 8;
+            for (var j: u32 = 0; j < 8; j++) {
+                let gw = iq1_grid[(ig + j) / 16];
+                let g = (gw >> (((ig + j) % 16) * 2)) & 3;
+                let gs = bitcast<i32>(g << 30) >> 30;
+                sum += dl * (f32(gs) + delta) * src1[src1_i];
+                src1_i++;
+            }
+        }
+    }
+    return sum;
+}
+
+#enddecl(IQ1_S)
+
+#decl(IQ1_M)
+
+struct iq1_m {
+    qs: array<u32, 8>,
+    qh: array<u32, 4>,
+    scales: array<u32, 2>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+
+    let scale = ((block.scales[0] >> 12) & 0xF) | ((block.scales[0] >> 24) & 0x00F0) | ((block.scales[1] >> 4) & 0x0F00) | ((block.scales[1] >> 16) & 0xF000);
+    let d = f32(bitcast<vec2<f16>>(scale).x);
+    var src1_i = src1_idx_base + offset * 256;
+    var sum = 0.0;
+    for (var ib: u32 = 0; ib < 8; ib++) {
+        let sw = (block.scales[ib / 4] >> (16 * ((ib / 2) % 2))) & 0xFFFF;
+        let s1 : u32 = (sw >> (6 * (ib % 2))) & 0x7;
+        let s2 : u32 = (sw >> (6 * (ib % 2) + 3)) & 0x7;
+        var dl = array<f32, 2>(
+            d * f32(2 * s1 + 1),
+            d * f32(2 * s2 + 1)
+        );
+
+        let qh = block.qh[ib / 2] >> (16 * (ib % 2));
+        var idx = array<u32, 4>(
+            get_byte(block.qs[ib], 0) | ((qh << 8) & 0x700),
+            get_byte(block.qs[ib], 1) | ((qh << 4) & 0x700),
+            get_byte(block.qs[ib], 2) | ((qh) & 0x700),
+            get_byte(block.qs[ib], 3) | ((qh >> 4) & 0x700)
+        );
+        var delta = array<f32, 4>(
+            select(IQ1_DELTA, -IQ1_DELTA, (qh & 0x08) != 0),
+            select(IQ1_DELTA, -IQ1_DELTA, (qh & 0x80) != 0),
+            select(IQ1_DELTA, -IQ1_DELTA, ((qh >> 8) & 0x08) != 0),
+            select(IQ1_DELTA, -IQ1_DELTA, ((qh >> 8) & 0x80) != 0)
+        );
+        for (var l: u32 = 0; l < 4; l++) {
+            let ig = idx[l] * 8;
+            for (var j: u32 = 0; j < 8; j++) {
+                let gw = iq1_grid[(ig + j) / 16];
+                let g = (gw >> (((ig + j) % 16) * 2)) & 3;
+                let gs = bitcast<i32>(g << 30) >> 30;
+                sum += dl[l/2] * (f32(gs) + delta[l]) * src1[src1_i];
+                src1_i++;
+            }
+        }
+    }
+    return sum;
+}
+
+#enddecl(IQ1_M)
+
+#decl(IQ4_TABLE)
+
+const kvalues_iq4nl = array<i32, 16>(
+    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113
+);
+
+#enddecl(IQ4_TABLE)
+
+#decl(IQ4_NL)
+
+struct iq4_nl {
+    d: f16,
+    qs: array<f16, 8>,
+}
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    var src1_i = src1_idx_base + offset * 32;
+    var sum = 0.0;
+    var qs: array<u32, 4>;
+    for (var i: u32 = 0; i < 4; i++) {
+        qs[i] = bitcast<u32>(vec2(block.qs[i * 2], block.qs[i * 2 + 1]));
+    }
+    for (var j: u32 = 0; j < 16; j++) {
+        let qsb = get_byte(qs[j / 4], j % 4);
+        sum += d * f32(kvalues_iq4nl[qsb & 0xF]) * src1[src1_i];
+        sum += d * f32(kvalues_iq4nl[qsb >> 4]) * src1[src1_i + 16];
+        src1_i++;
+    }
+    return sum;
+}
+
+#enddecl(IQ4_NL)
+
+#decl(IQ4_XS)
+
+struct iq4_xs {
+    d: f16,
+    scales_h: f16,
+    scales_l: u32,
+    qs: array<u32, 32>
+};
+
+fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
+    let block = src0[src0_idx_base + offset];
+    let d = f32(block.d);
+    let scales_h = bitcast<u32>(vec2(block.scales_h, 0.0));
+    var src1_i = src1_idx_base + offset * 256;
+    var sum = 0.0;
+    for (var ib: u32 = 0; ib < 8; ib++) {
+        let ls = ((get_byte(block.scales_l, ib / 2) >> (4 * (ib % 2))) & 0xF) | (((scales_h >> (2 * ib)) & 3) << 4);
+        let dl = d * (f32(ls) - 32.0);
+        for (var j: u32 = 0; j < 16; j++) {
+            let iqs = ib * 16 + j;
+            let qsb = get_byte(block.qs[iqs / 4], iqs % 4);
+            sum += dl * f32(kvalues_iq4nl[qsb & 0xF]) * src1[src1_i];
+            sum += dl * f32(kvalues_iq4nl[qsb >> 4]) * src1[src1_i + 16];
+            src1_i++;
+        }
+        src1_i += 16;
+    }
+    return sum;
+}
+
+#enddecl(IQ4_XS)
+
+#end(DECLS)
+
+#define(SHADER)
+
+enable f16;
+
+DECLS
+
+struct MulMatParams {
+    offset_src0: u32, // in elements/blocks
+    offset_src1: u32, // in elements/blocks
+    offset_dst: u32, // in elements/blocks
+    m: u32,
+    n: u32,
+    k: u32,
+    // all strides are in elements/blocks
+    stride_01: u32,
+    stride_11: u32,
+    stride_02: u32,
+    stride_12: u32,
+    stride_03: u32,
+    stride_13: u32,
+
+    bs02: u32,
+    bs03: u32,
+    broadcast2: u32,
+    broadcast3: u32
+};
+
+@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // N rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // M rows, K columns (transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns
+
+@group(0) @binding(3) var<uniform> params: MulMatParams;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
+    let total = params.m * params.n * params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
+    if (global_id.x >= total) {
+        return;
+    }
+
+    let dst2_stride = params.m * params.n;
+    let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
+
+    let dst3_idx = global_id.x / dst3_stride;
+    let src03_idx = dst3_idx / params.broadcast3; // src0 may be broadcast along the third dimension
+    let src13_idx = dst3_idx; // src1 is not broadcast
+    let dst3_rem = global_id.x % dst3_stride;
+
+    let dst2_idx = dst3_rem / dst2_stride;
+    let src02_idx = dst2_idx / params.broadcast2; // src0 may also be broadcast along the second dimension
+    let src12_idx = dst2_idx; // src1 is not broadcast
+
+    let dst2_rem = dst3_rem % dst2_stride;
+
+    let row = dst2_rem / params.n; // output row
+    let col = dst2_rem % params.n; // output column
+
+    let src0_idx_base = params.offset_src0 + src03_idx * params.stride_03 + src02_idx * params.stride_02 + col * params.stride_01;
+    let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11;
+
+    var sum = 0.0;
+    for (var i: u32 = 0u; i < params.k/{{BLOCK_SIZE}}; i = i + 1u) {
+        sum += multiply_add(src0_idx_base, src1_idx_base, i);
+    }
+    dst[params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.n + col] = sum;
+}
+
+#end(SHADER)
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl
deleted file mode 100644
index 054aab566f96b..0000000000000
--- a/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl
+++ /dev/null
@@ -1,56 +0,0 @@
-struct MulMatParams {
-    m: u32,
-    n: u32,
-    k: u32,
-    // all strides are in elements
-    stride_01: u32,
-    stride_11: u32,
-    stride_02: u32,
-    stride_12: u32,
-    stride_03: u32,
-    stride_13: u32,
-
-    bs02: u32,
-    bs03: u32,
-    broadcast2: u32,
-    broadcast3: u32
-};
-
-@group(0) @binding(0) var<storage, read_write> src0: array<f32>; // N rows, K columns
-@group(0) @binding(1) var<storage, read_write> src1: array<f32>; // M rows, K columns (transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns
-
-@group(0) @binding(3) var<uniform> params: MulMatParams;
-
-@compute @workgroup_size(64)
-fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
-    let total = params.m * params.n * params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
-    if (global_id.x >= total) {
-        return;
-    }
-
-    let dst2_stride = params.m * params.n;
-    let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
-
-    let dst3_idx = global_id.x / dst3_stride;
-    let src03_idx = dst3_idx / params.broadcast3; // src0 may be broadcast along the third dimension
-    let src13_idx = dst3_idx; // src1 is not broadcast
-    let dst3_rem = global_id.x % dst3_stride;
-
-    let dst2_idx = dst3_rem / dst2_stride;
-    let src02_idx = dst2_idx / params.broadcast2; // src0 may also be broadcast along the second dimension
-    let src12_idx = dst2_idx; // src1 is not broadcast
-
-    let dst2_rem = dst3_rem % dst2_stride;
-
-    let row = dst2_rem / params.n; // output row
-    let col = dst2_rem % params.n; // output column
-
-    var sum = 0.0;
-    for (var i: u32 = 0u; i < params.k; i = i + 1u) {
-        let src0_idx = src03_idx * params.stride_03 + src02_idx * params.stride_02 + col * params.stride_01 + i;
-        let src1_idx = src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11 + i;
-        sum = sum + src0[src0_idx] * src1[src1_idx];
-    }
-    dst[dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.n + col] = sum;
-}
diff --git a/ggml/src/ggml-zdnn/CMakeLists.txt b/ggml/src/ggml-zdnn/CMakeLists.txt
new file mode 100644
index 0000000000000..0a723ce4de286
--- /dev/null
+++ b/ggml/src/ggml-zdnn/CMakeLists.txt
@@ -0,0 +1,36 @@
+if (DEFINED ZDNN_ROOT)
+    message(STATUS "zdnn: using ZDNN_ROOT override: ${ZDNN_ROOT}")
+    set(ZDNN_HINT "${ZDNN_ROOT}")
+else()
+    set(ZDNN_HINT "")
+endif()
+
+find_path(ZDNN_INCLUDE
+            NAMES zdnn.h
+            HINTS ${ZDNN_HINT} /usr /usr/local
+            PATH_SUFFIXES include)
+if (ZDNN_INCLUDE)
+    message(STATUS "zdnn: found include: ${ZDNN_INCLUDE}")
+else()
+    message(FATAL_ERROR "zdnn: include directory not found, please set ZDNN_ROOT to the proper path if necessary")
+endif()
+
+find_library(ZDNN_LIB
+                NAMES zdnn
+                HINTS ${ZDNN_HINT} /usr /usr/local
+                PATH_SUFFIXES lib lib64)
+if (ZDNN_LIB)
+    message(STATUS "zdnn: found library: ${ZDNN_LIB}")
+else()
+    message(FATAL_ERROR "zdnn: library not found, please set ZDNN_ROOT to the proper path if necessary")
+endif()
+
+file(GLOB GGML_SOURCES_ZDNN "*.c" "*.cpp")
+file(GLOB GGML_HEADERS_ZDNN "*.h" "*.hpp")
+
+ggml_add_backend_library(ggml-zdnn ${GGML_HEADERS_ZDNN} ${GGML_SOURCES_ZDNN})
+target_link_libraries(ggml-zdnn PRIVATE ${ZDNN_LIB})
+target_include_directories(ggml-zdnn PRIVATE ${ZDNN_INCLUDE})
+target_link_directories(ggml-zdnn PRIVATE ${ZDNN_LIB})
+
+target_compile_definitions(ggml-zdnn PRIVATE GGML_USE_ZDNN)
diff --git a/ggml/src/ggml-zdnn/ggml-zdnn-impl.h b/ggml/src/ggml-zdnn/ggml-zdnn-impl.h
new file mode 100644
index 0000000000000..9dcb040fa894e
--- /dev/null
+++ b/ggml/src/ggml-zdnn/ggml-zdnn-impl.h
@@ -0,0 +1,97 @@
+#ifndef GGML_ZDNN_IMPL
+#define GGML_ZDNN_IMPL
+
+#include "zdnn.h"
+#include "ggml.h"
+#include "ggml-zdnn.h"
+
+#include <vector>
+#include <memory>
+#include <vecintrin.h>
+
+#define GGML_ZDNN_NAME    "zDNN"
+#define GGML_ZDNN_VERSION ZDNN_VERNUM
+
+#define vec_neg(a)    (-(a))                // Vector Negate
+#define vec_add(a, b) ((a) + (b))           // Vector Add
+#define vec_sub(a, b) ((a) - (b))           // Vector Subtract
+#define vec_mul(a, b) ((a) * (b))           // Vector Multiply
+#define vec_div(a, b) ((a) / (b))           // Vector Divide
+#define vec_sl(a, b)  ((a) << (b))          // Vector Shift Left
+#define vec_sra(a, b) ((a) >> (b))          // Vector Shift Right
+#define vec_sr(a, b)  ((a) >> (b))          // Vector Shift Right Algebraic
+#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
+#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
+
+#ifndef vec_and
+#define vec_and(a, b) ((a) & (b)) // Vector AND
+#endif
+
+#ifndef vec_or
+#define vec_or(a, b)  ((a) | (b)) // Vector OR
+#endif
+
+#ifndef vec_xor
+#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
+#endif
+
+typedef   signed char char8x16_t  __attribute__((vector_size(16)));
+typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
+
+typedef int8_t   int8x16_t  __attribute__((vector_size(16)));
+typedef int16_t  int16x8_t  __attribute__((vector_size(16)));
+typedef int32_t  int32x4_t  __attribute__((vector_size(16)));
+typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
+typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
+typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
+
+typedef float float32x4_t   __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute__((vector_size(16)));
+
+typedef   signed long long long64x2_t  __attribute__((vector_size(16)));
+typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
+
+#define ZDNN_CHECK(stmt)                \
+    do {                                \
+        zdnn_status status = (stmt);    \
+        GGML_ASSERT(status == ZDNN_OK); \
+    } while (0);
+
+struct ggml_backend_zdnn_device_context {
+    int zdnn_device;
+    int zdnn_device_ref_count;
+
+    bool has_parmblkformat_0;
+    bool has_parmblkformat_1;
+
+    size_t max_size;
+
+    char name[128];
+};
+
+struct ggml_backend_zdnn_context {
+    int device;
+    ggml_cgraph * gf;
+};
+
+struct ggml_backend_zdnn_buffer {
+    void * data;
+    size_t size;
+
+    zdnn_tensor_desc pre_tfm_desc;
+    zdnn_tensor_desc tfm_desc;
+    zdnn_ztensor     ztensor;
+
+    char name[GGML_MAX_NAME];
+};
+
+struct ggml_backend_zdnn_buffer_context {
+    void * all_data;
+    size_t all_size;
+    bool owned;
+
+    int n_buffers;
+    std::vector<std::unique_ptr<ggml_backend_zdnn_buffer>> buffers;
+};
+
+#endif  // GGML_ZDNN_IMPL
diff --git a/ggml/src/ggml-zdnn/ggml-zdnn.cpp b/ggml/src/ggml-zdnn/ggml-zdnn.cpp
new file mode 100644
index 0000000000000..7507a52aeae31
--- /dev/null
+++ b/ggml/src/ggml-zdnn/ggml-zdnn.cpp
@@ -0,0 +1,846 @@
+#include "zdnn.h"
+#include "ggml-zdnn.h"
+#include "ggml-zdnn-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include <vector>
+#include <memory>
+#include <csignal>
+#include <unistd.h>
+
+inline zdnn_data_types ggml_zdnn_type_mapping(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return FP32;
+        case GGML_TYPE_F16:
+            return FP16;
+        case GGML_TYPE_BF16:
+            return BFLOAT;
+        case GGML_TYPE_I8:
+            return INT8;
+        case GGML_TYPE_I32:
+            return INT32;
+        case GGML_TYPE_Q8_0:
+            return INT8;
+        default:
+            GGML_ABORT("%s: fatal: unable to determine zTensor data type",
+                       __func__);
+            break;
+    }
+}
+
+inline void ggml_zdnn_create_tensor(zdnn_tensor_desc  & pre_tfm_desc,
+                                    zdnn_tensor_desc  & tfm_desc,
+                                    zdnn_ztensor      & ztensor,
+                              const ggml_tensor       * src,
+                              const int64_t           * ne,
+                              const zdnn_data_layouts   layout) {
+    zdnn_init_pre_transformed_desc(
+        layout,
+        ggml_zdnn_type_mapping(src->type),
+        &pre_tfm_desc,
+        ne[3], ne[2], ne[1], ne[0]
+    );
+
+    ZDNN_CHECK(zdnn_generate_transformed_desc(&pre_tfm_desc, &tfm_desc));
+    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&pre_tfm_desc, &tfm_desc, &ztensor));
+}
+
+inline void ggml_zdnn_load_tensor(zdnn_ztensor & ztensor,
+                                          void * buffer) {
+    ZDNN_CHECK(zdnn_transform_ztensor(&ztensor, buffer));
+}
+
+inline void ggml_zdnn_init_tensor(ggml_backend_zdnn_buffer * buffer, const ggml_tensor * tensor) {
+    switch (tensor->op) {
+        case GGML_OP_MUL_MAT:
+            {
+                zdnn_init_pre_transformed_desc(
+                    ZDNN_2D,
+                    ggml_zdnn_type_mapping(tensor->type),
+                    &buffer->pre_tfm_desc,
+                    tensor->ne[1], tensor->ne[0]
+                );
+            } break;
+
+        default:
+            {
+                // For 4D tensors, GGML uses NCHW layout. However, because zDNN
+                // automatically transforms everything to NHWC, we will use it
+                // directly to avoid the performance penalty changing the
+                // layout and reshaping the tensor.
+                zdnn_init_pre_transformed_desc(
+                    ZDNN_NHWC,
+                    ggml_zdnn_type_mapping(tensor->type),
+                    &buffer->pre_tfm_desc,
+                    tensor->ne[3], tensor->ne[2], tensor->ne[1], tensor->ne[0]
+                );
+
+                // TODO: Consider adding a ggml check.
+                // TODO: If tensor = 4D, use ZDNN_NCHW by default.
+                // TODO: If tensor = 2D, use ZDNN_NHWC by default.
+            } break;
+    }
+
+    ZDNN_CHECK(zdnn_generate_transformed_desc(&buffer->pre_tfm_desc, &buffer->tfm_desc));
+    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&buffer->pre_tfm_desc, &buffer->tfm_desc, &buffer->ztensor));
+}
+
+static void ggml_zdnn_mul_mat_op(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    const ggml_tensor * weights = src0;
+    const ggml_tensor * inputs  = src1;
+          ggml_tensor * output  = dst;
+
+    ggml_backend_zdnn_buffer * weights_extra = (ggml_backend_zdnn_buffer *)weights->extra;
+    ggml_backend_zdnn_buffer * inputs_extra  = (ggml_backend_zdnn_buffer *)inputs->extra;
+    ggml_backend_zdnn_buffer * output_extra  = (ggml_backend_zdnn_buffer *)output->extra;
+
+    zdnn_tensor_desc ptd_bias, td_bias;
+    zdnn_ztensor zt_bias;
+
+    const int64_t weights_rows = ne01;
+    const int64_t weights_cols = ne00;
+    const int64_t inputs_rows  = ne11;
+    const int64_t inputs_cols  = ne10;
+
+    assert(inputs_cols == weights_cols);
+
+    const int64_t output_rows = ne1;
+    const int64_t output_cols = ne0;
+
+    const int64_t bias_dim  [GGML_MAX_DIMS]  = { 1, 1, 1, output_cols };
+    ggml_zdnn_create_tensor(ptd_bias, td_bias, zt_bias, output, bias_dim, ZDNN_1D);
+
+    void * bias_data = (void *)calloc(ne0, ggml_element_size(output));
+    if (weights_extra->ztensor.is_transformed == false) ggml_zdnn_load_tensor(weights_extra->ztensor, weights->data);
+    if (inputs_extra->ztensor.is_transformed == false) ggml_zdnn_load_tensor(inputs_extra->ztensor, inputs->data);
+    ggml_zdnn_load_tensor(zt_bias, bias_data);
+
+    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
+    //               __func__, weights_extra->name,
+    //               weights->ne[3], weights->ne[2], weights->ne[1], weights->ne[0],
+    //               weights_extra->pre_tfm_desc.dim1,
+    //               weights_extra->pre_tfm_desc.dim2,
+    //               weights_extra->pre_tfm_desc.dim3,
+    //               weights_extra->pre_tfm_desc.dim4);
+
+    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
+    //               __func__, inputs_extra->name,
+    //               inputs->ne[3], inputs->ne[2], inputs->ne[1], inputs->ne[0],
+    //               inputs_extra->pre_tfm_desc.dim1,
+    //               inputs_extra->pre_tfm_desc.dim2,
+    //               inputs_extra->pre_tfm_desc.dim3,
+    //               inputs_extra->pre_tfm_desc.dim4);
+
+    GGML_ASSERT(weights_extra->pre_tfm_desc.dim1 == weights->ne[0] && "weights_extra->pre_tfm_desc.dim1 must match weights->ne[0]");
+    GGML_ASSERT(weights_extra->pre_tfm_desc.dim2 == weights->ne[1] && "weights_extra->pre_tfm_desc.dim2 must match weights->ne[1]");
+    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim1  == inputs->ne[0]  && "inputs_extra->pre_tfm_desc.dim1 must match inputs->ne[0]");
+    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim2  == inputs->ne[1]  && "inputs_extra->pre_tfm_desc.dim2 must match inputs->ne[1]");
+
+    ZDNN_CHECK(zdnn_matmul_transpose_op(&inputs_extra->ztensor, &weights_extra->ztensor, &zt_bias,
+                                        false, true, MATMUL_OP_ADDITION, &output_extra->ztensor));
+    // TODO: Remove in the future as we are currently DLF16 -> FP32 then in the next op, FP32 -> DLF16 again. Inefficient.
+    ZDNN_CHECK(zdnn_transform_origtensor(&output_extra->ztensor, output->data));
+
+    ZDNN_CHECK(zdnn_free_ztensor_buffer(&zt_bias));
+    free(bias_data);
+}
+
+static void ggml_zdnn_mul_mat_dispatch(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    bool use_mul_mat_vec =
+        (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F16)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+
+    bool use_mul_mat_vec_q =
+        ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    bool use_mul_mat_q =
+        ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    // debug helpers
+    // GGML_LOG_INFO("%s: use_mul_mat_vec   = %d\n", __func__, use_mul_mat_vec);
+    // GGML_LOG_INFO("%s: use_mul_mat_vec_q = %d\n", __func__, use_mul_mat_vec_q);
+    // GGML_LOG_INFO("%s: use_mul_mat_q     = %d\n", __func__, use_mul_mat_q);
+    // GGML_LOG_INFO("%s: src0: %8d %8d %8d %8d\n", __func__, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
+    // GGML_LOG_INFO("%s:       %8d %8d %8d %8d\n", __func__, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
+    // GGML_LOG_INFO("%s: src1: %8d %8d %8d %8d\n", __func__, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
+    // GGML_LOG_INFO("%s:       %8d %8d %8d %8d\n", __func__, src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
+    // GGML_LOG_INFO("%s: src0 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
+    // GGML_LOG_INFO("%s: src1 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
+
+    if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16
+        && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)
+        && src1->ne[2] * src1->ne[3] > 1) {
+        // general KQ + KQV multi-batch
+        GGML_LOG_INFO("%s: using zdnn_mul_mat_batched for KQ + KQV multi-batch\n", __func__);
+        // ggml_zdnn_mul_mat_batched(ctx, src0, src1, dst);
+    } else if (use_mul_mat_vec) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_vec for vector multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec_q) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_vec_q for quantized vector multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_vec_q, ggml_zdnn_quantize_row_q8_1);
+    } else if (use_mul_mat_q) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_q for quantized matrix multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_q, ggml_zdnn_quantize_mmq_q8_1);
+    } else {
+        // GGML_LOG_INFO("%s: using zdnn_op_mul_mat for general matrix multiplication\n", __func__);
+        ggml_zdnn_mul_mat_op(ctx, src0, src1, dst);
+    }
+}
+
+static bool ggml_zdnn_compute_forward(ggml_backend_zdnn_context * ctx, ggml_tensor * dst) {
+    switch (dst->op) {
+        case GGML_OP_MUL_MAT:
+            ggml_zdnn_mul_mat_dispatch(ctx, dst->src[0], dst->src[1], dst);
+            break;
+
+        default:
+            return false;
+    }
+
+    return true;
+}
+
+static enum ggml_status ggml_zdnn_graph_compute(ggml_backend_t backend, ggml_cgraph * gf) {
+    ggml_backend_zdnn_context        * ctx     = (       ggml_backend_zdnn_context *)backend->context;
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)backend->device->context;
+
+    ctx->gf = gf;
+    for (int i = 0; i < gf->n_nodes; i++) {
+        ggml_tensor * node = gf->nodes[i];
+
+        if (ggml_is_empty(node)
+            || node->op == GGML_OP_NONE
+            || node->op == GGML_OP_RESHAPE
+            || node->op == GGML_OP_VIEW
+            || node->op == GGML_OP_PERMUTE
+            || node->op == GGML_OP_TRANSPOSE) {
+            continue;
+        }
+
+        bool ok = ggml_zdnn_compute_forward(ctx, node);
+        if (!ok) {
+            GGML_LOG_ERROR("%s: unsupported op %s (%s)\n",
+                           __func__, node->name, ggml_op_name(node->op));
+        }
+
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static bool ggml_zdnn_supports_op(const ggml_backend_zdnn_device_context * ctx_dev, const ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_PERMUTE:
+            return true;
+
+        case GGML_OP_MUL_MAT:
+            {
+                const ggml_tensor * src0 = op->src[0];
+                const ggml_tensor * src1 = op->src[1];
+
+                const int64_t ne10 = src1->ne[0];
+                const int64_t ne0 = op->ne[0];
+                const int64_t ne1 = op->ne[1];
+
+                const int64_t max_batch = ctx_dev->max_size;
+
+                return ggml_is_matrix(src0) &&
+                       ggml_is_matrix(src1) &&
+                       ggml_is_contiguous(src0) &&
+                       ggml_is_contiguous(src1) &&
+                       src0->view_src == nullptr && src1->view_src == nullptr &&
+                       src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 &&
+                       (ne0 <= max_batch && ne1 <= max_batch && ne10 <= max_batch);
+            } break;
+
+        default:
+            return false;
+    }
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+//
+// globals
+//
+
+// initialised in ggml_backend_zdnn_reg
+static ggml_backend_reg    g_ggml_backend_zdnn_reg;
+static ggml_backend_device g_ggml_backend_zdnn_device;
+
+static ggml_backend_zdnn_device_context g_ggml_ctx_dev_main = {
+    /* .zdnn_device           = */ 0,
+    /* .zdnn_device_ref_count = */ 0,
+    /* .has_parmblkformat_0   = */ false,
+    /* .has_parmblkformat_1   = */ false,
+    /* .max_size              = */ 0,
+    /* .name                  = */ "",
+};
+
+static int ggml_backend_zdnn_device_acq(ggml_backend_zdnn_device_context * ctx) {
+    assert(ctx != NULL);
+
+    if (ctx->zdnn_device == 0) {
+        ctx->zdnn_device = 1;
+    }
+
+    if (ctx->zdnn_device >= 1) {
+        ctx->has_parmblkformat_0 = zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_0);
+        ctx->has_parmblkformat_1 = zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_1);
+        ctx->max_size = zdnn_get_nnpa_max_dim_idx_size();
+        strncpy(ctx->name, GGML_ZDNN_NAME, sizeof(ctx->name) - 1);
+    }
+
+    ctx->zdnn_device_ref_count++;
+    return ctx->zdnn_device;
+}
+
+static void ggml_backend_zdnn_device_rel(ggml_backend_zdnn_device_context * ctx) {
+    assert(ctx != NULL);
+    assert(ctx->zdnn_device_ref_count > 0);
+
+    ctx->zdnn_device_ref_count--;
+    if (ctx->zdnn_device_ref_count == 0) {
+        if (ctx->zdnn_device >= 0) {
+            ctx->zdnn_device = 0;
+        }
+    }
+}
+
+static ggml_backend_zdnn_context * ggml_zdnn_init(ggml_backend_dev_t dev) {
+    GGML_LOG_INFO("%s: allocating\n", __func__);
+    GGML_LOG_INFO("%s: found 1 device\n", __func__);
+
+    #ifdef STATIC_LIB
+    zdnn_init();
+    #endif
+
+    ggml_backend_zdnn_context * ctx = new ggml_backend_zdnn_context();
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)dev->context;
+
+    int device = 1;
+    GGML_LOG_INFO("%s: picking default device: %s\n", __func__, ctx_dev->name);
+
+    ctx->device = device;
+    GGML_LOG_INFO("%s: NNPA name: %s\n", __func__, ctx_dev->name);
+    GGML_LOG_INFO("%s: NNPA_PARMBLKFORMAT_0 = %s\n", __func__, ctx_dev->has_parmblkformat_0 ? "true" : "false");
+    GGML_LOG_INFO("%s: NNPA_PARMBLKFORMAT_1 = %s\n", __func__, ctx_dev->has_parmblkformat_1 ? "true" : "false");
+
+    ctx->gf = nullptr;
+
+    return ctx;
+}
+
+static void ggml_zdnn_free(ggml_backend_zdnn_context * ctx) {
+    GGML_LOG_INFO("%s: deallocating\n", __func__);
+    delete ctx;
+}
+
+//
+// backend interface
+//
+
+static void ggml_backend_zdnn_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    for (int i = 0; i < ctx->n_buffers; i++) {
+        if (ctx->buffers[i]->ztensor.buffer != NULL && ctx->buffers[i]->ztensor.is_transformed) {
+            ZDNN_CHECK(zdnn_free_ztensor_buffer(&ctx->buffers[i]->ztensor));
+        }
+    }
+
+    delete ctx;
+}
+
+static void * ggml_backend_zdnn_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+    return ctx->all_data;
+}
+
+static enum ggml_status ggml_backend_zdnn_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    if (tensor->view_src != NULL) {
+        assert(tensor->view_src->buffer->buft == buffer->buft);
+        return GGML_STATUS_SUCCESS;
+    }
+
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    const int64_t tsize = ggml_nbytes(tensor);
+    int buffer_idx = ctx->n_buffers;
+
+    std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+    zdnn_buffer->data = tensor->data;
+    zdnn_buffer->size = tsize;
+    strncpy(zdnn_buffer->name, tensor->name, GGML_MAX_NAME - 1);
+
+    ggml_zdnn_init_tensor(zdnn_buffer.get(), tensor);
+    tensor->extra = zdnn_buffer.get();
+
+    ctx->buffers.push_back(std::move(zdnn_buffer));
+    ctx->n_buffers++;
+
+    // GGML_LOG_INFO("%s: initialised tensor '%s' in buffer %d, size = %8.2f MiB\n",
+    //               __func__, tensor->name, buffer_idx, tsize);
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static void ggml_backend_zdnn_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    memset(ctx->all_data, value, ctx->all_size);
+}
+
+static ggml_backend_buffer_i ggml_backend_zdnn_buffer_i = {
+    /* .free_buffer   = */ ggml_backend_zdnn_buffer_free_buffer,
+    /* .get_base      = */ ggml_backend_zdnn_buffer_get_base,
+    /* .init_tensor   = */ ggml_backend_zdnn_buffer_init_tensor,
+    /* .memset_tensor = */ ggml_backend_zdnn_buffer_memset_tensor,
+    /* .set_tensor    = */ ggml_backend_zdnn_buffer_set_tensor,
+    /* .get_tensor    = */ ggml_backend_zdnn_buffer_get_tensor,
+    /* .cpy_tensor    = */ NULL,
+    /* .clear         = */ ggml_backend_zdnn_buffer_clear,
+    /* .reset         = */ NULL,
+};
+
+//
+// default buffer type
+//
+
+static const char * ggml_backend_zdnn_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_zdnn_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    ggml_backend_zdnn_buffer_context * ctx = new ggml_backend_zdnn_buffer_context();
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += size_page - (size_aligned % size_page);
+    }
+
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)buft->device->context;
+
+    GGML_ASSERT(ctx_dev->zdnn_device >= 0);
+    int device = ctx_dev->zdnn_device; GGML_UNUSED(device);
+
+    ctx->all_data  = ggml_aligned_malloc(size_aligned);
+    ctx->all_size  = size_aligned;
+    ctx->owned     = true;
+    ctx->n_buffers = 1;
+
+    if (ctx->all_data != NULL) {
+        std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+        zdnn_buffer->data = ctx->all_data;
+        zdnn_buffer->size = size_aligned;
+        ctx->buffers.push_back(std::move(zdnn_buffer));
+    }
+
+    if (size_aligned > 0 && (ctx->all_data == NULL)) {
+        GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f\n",
+                       __func__, size_aligned / 1024.0 / 1024.0);
+        delete ctx;
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_zdnn_buffer_i, ctx, size);
+}
+
+static size_t ggml_backend_zdnn_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 256;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_zdnn_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_zdnn_buffer_type(void) {
+    static ggml_backend_buffer_type ggml_backend_buffer_type_zdnn = {
+        /* .iface   = */ {
+            /* .get_name       = */ ggml_backend_zdnn_buffer_type_get_name,
+            /* .alloc_buffer   = */ ggml_backend_zdnn_buffer_type_alloc_buffer,
+            /* .get_alignment  = */ ggml_backend_zdnn_buffer_type_get_alignment,
+            /* .get_max_size   = */ NULL,
+            /* .get_alloc_size = */ NULL,  // defaults to ggml_nbytes
+            /* .is_host        = */ ggml_backend_zdnn_buffer_type_is_host,
+        },
+        /* .device  = */ &g_ggml_backend_zdnn_device,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_type_zdnn;
+}
+
+static const char * ggml_backend_zdnn_buffer_from_ptr_type_get_name(ggml_backend_buffer_type_t buft) {
+    return GGML_ZDNN_NAME "_Mapped";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_zdnn_buffer_from_ptr_type(void) {
+    static ggml_backend_buffer_type ggml_backend_buffer_from_ptr_type_zdnn = {
+        /* .iface = */ {
+            /* .get_name       = */ ggml_backend_zdnn_buffer_from_ptr_type_get_name,
+            /* .alloc_buffer   = */ ggml_backend_zdnn_buffer_type_alloc_buffer,
+            /* .get_alignment  = */ ggml_backend_zdnn_buffer_type_get_alignment,
+            /* .get_max_size   = */ NULL,
+            /* .get_alloc_size = */ NULL,  // defaults to ggml_nbytes
+            /* .is_host        = */ ggml_backend_zdnn_buffer_type_is_host,
+        },
+        /* .device  = */ &g_ggml_backend_zdnn_device,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_from_ptr_type_zdnn;
+}
+
+//
+// backend
+//
+
+static const char * ggml_backend_zdnn_name(ggml_backend_t backend) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_zdnn_free(ggml_backend_t backend) {
+    ggml_backend_zdnn_context * ctx = (ggml_backend_zdnn_context *)backend->context;
+
+    ggml_zdnn_free(ctx);
+    free(backend);
+}
+
+static enum ggml_status ggml_backend_zdnn_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    return ggml_zdnn_graph_compute(backend, cgraph);
+}
+
+static ggml_backend_i ggml_backend_zdnn_i = {
+    /* .get_name           = */ ggml_backend_zdnn_name,
+    /* .free               = */ ggml_backend_zdnn_free,
+    /* .set_tensor_async   = */ NULL,
+    /* .get_tensor_async   = */ NULL,
+    /* .cpy_tensor_async   = */ NULL,
+    /* .synchronize        = */ NULL,
+    /* .graph_plan_create  = */ NULL,
+    /* .graph_plan_free    = */ NULL,
+    /* .graph_plan_update  = */ NULL,
+    /* .graph_plan_compute = */ NULL,
+    /* .graph_compute      = */ ggml_backend_zdnn_graph_compute,
+    /* .event_record       = */ NULL,
+    /* .event_wait         = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_zdnn_guid(void) {
+    static const char * guid_str = "IBM-ZDNN-ACCELER";
+    return reinterpret_cast<ggml_guid_t>((void *)guid_str);
+}
+
+// TODO: remove in the future
+ggml_backend_t ggml_backend_zdnn_init(void) {
+    ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_zdnn_reg(), 0);
+
+    ggml_backend_zdnn_context * ctx = ggml_zdnn_init(dev);
+    if (ctx == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_t backend = (ggml_backend_t)malloc(sizeof(ggml_backend));
+    *backend = (ggml_backend) {
+        /* .guid       = */ ggml_backend_zdnn_guid(),
+        /* .iface      = */ ggml_backend_zdnn_i,
+        /* .device     = */ dev,
+        /* .context    = */ ctx,
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_zdnn(ggml_backend_t backend) {
+    return backend != NULL &&
+           ggml_guid_matches(backend->guid, ggml_backend_zdnn_guid());
+
+    GGML_UNUSED(backend);
+}
+
+//
+// backend device
+//
+
+static const char * ggml_backend_zdnn_device_get_name(ggml_backend_dev_t dev) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_zdnn_device_get_description(ggml_backend_dev_t dev) {
+    return "IBM Z Neural Network Processing Assist (NNPA)";
+}
+
+static void ggml_backend_zdnn_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    *free  = 0;
+    *total = 0;
+}
+
+static enum ggml_backend_dev_type ggml_backend_zdnn_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_zdnn_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_zdnn_device_get_name(dev);
+    props->description = ggml_backend_zdnn_device_get_description(dev);
+    props->type        = ggml_backend_zdnn_device_get_type(dev);
+    ggml_backend_zdnn_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = (ggml_backend_dev_caps) {
+        /* .async                = */ false,
+        /* .host_buffer          = */ false,
+        /* .buffer_from_host_ptr = */ true,
+        /* .events               = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_zdnn_device_init(ggml_backend_dev_t dev, const char * params) {
+    ggml_backend_zdnn_context * ctx = ggml_zdnn_init(dev);
+    if (ctx == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_t backend = (ggml_backend *)malloc(sizeof(ggml_backend));
+    *backend = (ggml_backend) {
+        /* .guid       = */ ggml_backend_zdnn_guid(),
+        /* .iface      = */ ggml_backend_zdnn_i,
+        /* .device     = */ dev,
+        /* .context    = */ ctx,
+    };
+
+    return backend;
+
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_zdnn_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_zdnn_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_zdnn_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    ggml_backend_zdnn_buffer_context * ctx = new ggml_backend_zdnn_buffer_context();
+
+    ctx->all_data  = ptr;
+    ctx->all_size  = size;
+    ctx->owned     = false;
+    ctx->n_buffers = 0;
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    // page-align the data ptr
+    {
+        const uintptr_t offs = (uintptr_t) ptr % size_page;
+        ptr  = (void *)((char *)ptr - offs);
+        size += offs;
+    }
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += size_page - (size_aligned % size_page);
+    }
+
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)dev->context;
+
+    GGML_ASSERT(ctx_dev->zdnn_device >= 0);
+    int device = ctx_dev->zdnn_device; GGML_UNUSED(device);
+
+    std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+    zdnn_buffer->data = ptr;
+    zdnn_buffer->size = size;
+    ctx->buffers.push_back(std::move(zdnn_buffer));
+
+    GGML_LOG_INFO("%s: allocated buffer, size = %8.2f MiB\n",
+                  __func__, size_aligned / 1024.0 / 1024.0);
+
+    ++ctx->n_buffers;
+
+    return ggml_backend_buffer_init(ggml_backend_zdnn_buffer_from_ptr_type(), ggml_backend_zdnn_buffer_i, ctx, size);
+}
+
+static bool ggml_backend_zdnn_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *) dev->context;
+
+    return ggml_zdnn_supports_op(ctx_dev, op);
+}
+
+static bool ggml_backend_zdnn_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return
+        buft->iface.get_name == ggml_backend_zdnn_buffer_type_get_name ||
+        buft->iface.get_name == ggml_backend_zdnn_buffer_from_ptr_type_get_name;
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_device_i ggml_backend_zdnn_device_i = {
+    /* .get_name             = */ ggml_backend_zdnn_device_get_name,
+    /* .get_description      = */ ggml_backend_zdnn_device_get_description,
+    /* .get_memory           = */ ggml_backend_zdnn_device_get_memory,
+    /* .get_type             = */ ggml_backend_zdnn_device_get_type,
+    /* .get_props            = */ ggml_backend_zdnn_device_get_props,
+    /* .init_backend         = */ ggml_backend_zdnn_device_init,
+    /* .get_buffer_type      = */ ggml_backend_zdnn_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_zdnn_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_zdnn_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_zdnn_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+//
+// backend registry
+//
+
+static const char * ggml_backend_zdnn_reg_get_name(ggml_backend_reg_t reg) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_zdnn_reg_device_count(ggml_backend_reg_t reg) {
+    if (!zdnn_is_nnpa_installed()) {
+        return 0;
+    }
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_zdnn_reg_device_get(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    return &g_ggml_backend_zdnn_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static ggml_backend_feature g_ggml_backend_zdnn_features[] = {
+    { "NNPA", zdnn_is_nnpa_installed() ? "1" : "0" },
+    { "NNPA_PARMBLKFORMAT_0", zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_0) ? "1" : "0" },
+    { "NNPA_PARMBLKFORMAT_1", zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_1) ? "1" : "0" },
+    { NULL, NULL },
+};
+
+static ggml_backend_feature * ggml_backend_zdnn_get_features(ggml_backend_reg_t reg) {
+    return g_ggml_backend_zdnn_features;
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_zdnn_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *) ggml_backend_zdnn_get_features;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_reg_i ggml_backend_zdnn_reg_i = {
+    /* .get_name         = */ ggml_backend_zdnn_reg_get_name,
+    /* .get_device_count = */ ggml_backend_zdnn_reg_device_count,
+    /* .get_device       = */ ggml_backend_zdnn_reg_device_get,
+    /* .get_proc_address = */ ggml_backend_zdnn_get_proc_address,
+};
+
+static void ggml_zdnn_cleanup(void) {
+    ggml_backend_zdnn_device_rel(&g_ggml_ctx_dev_main);
+}
+
+// TODO: make thread-safe
+ggml_backend_reg_t ggml_backend_zdnn_reg(void) {
+    ggml_backend_zdnn_device_acq(&g_ggml_ctx_dev_main);
+
+    // register cleanup callback
+    atexit(ggml_zdnn_cleanup);
+
+    {
+        g_ggml_backend_zdnn_reg = (ggml_backend_reg) {
+            /* .api_version = */ GGML_ZDNN_VERSION,
+            /* .iface       = */ ggml_backend_zdnn_reg_i,
+            /* .context     = */ NULL,
+        };
+
+        g_ggml_backend_zdnn_device = (ggml_backend_device) {
+            /* .iface       = */ ggml_backend_zdnn_device_i,
+            /* .reg         = */ &g_ggml_backend_zdnn_reg,
+            /* .context     = */ &g_ggml_ctx_dev_main,
+        };
+
+        return &g_ggml_backend_zdnn_reg;
+    }
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_zdnn_reg)
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 55a76f8248c09..d76ea58f789e2 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -975,6 +975,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "IM2COL",
     "IM2COL_BACK",
     "CONV_2D",
+    "CONV_3D",
     "CONV_2D_DW",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
@@ -1012,11 +1013,12 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS",
     "CROSS_ENTROPY_LOSS_BACK",
     "OPT_STEP_ADAMW",
+    "OPT_STEP_SGD",
 
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 87, "GGML_OP_COUNT != 87");
+static_assert(GGML_OP_COUNT == 89, "GGML_OP_COUNT != 89");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1076,6 +1078,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "im2col(x)",
     "im2col_back(x)",
     "conv_2d(x)",
+    "conv_3d(x)",
     "conv_2d_dw(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
@@ -1113,15 +1116,15 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss(x,y)",
     "cross_entropy_loss_back(x,y)",
     "adamw(x)",
+    "sgd(x)",
 
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 87, "GGML_OP_COUNT != 87");
+static_assert(GGML_OP_COUNT == 89, "GGML_OP_COUNT != 89");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
-
 static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "ABS",
     "SGN",
@@ -3885,6 +3888,7 @@ static struct ggml_tensor * ggml_rope_impl(
         struct ggml_tensor  * b,
         struct ggml_tensor  * c,
         int                   n_dims,
+        int                   sections[GGML_MROPE_SECTIONS],
         int                   mode,
         int                   n_ctx_orig,
         float                 freq_base,
@@ -3898,15 +3902,19 @@ static struct ggml_tensor * ggml_rope_impl(
 
     GGML_ASSERT(ggml_is_vector(b));
     GGML_ASSERT(b->type == GGML_TYPE_I32);
-    GGML_ASSERT(a->ne[2] == b->ne[0]);
+
+    bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;
+    if (mrope_used) {
+        GGML_ASSERT(a->ne[2] * 4 == b->ne[0]); // mrope expecting 4 position ids per token
+    } else {
+        GGML_ASSERT(a->ne[2] == b->ne[0]);
+    }
 
     if (c) {
         GGML_ASSERT(c->type == GGML_TYPE_F32);
         GGML_ASSERT(c->ne[0] >= n_dims / 2);
     }
 
-    int sections[4] = {0, 0, 0, 0};
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[15] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
@@ -3916,7 +3924,11 @@ static struct ggml_tensor * ggml_rope_impl(
     memcpy(params +  8, &attn_factor,  sizeof(float));
     memcpy(params +  9, &beta_fast,    sizeof(float));
     memcpy(params + 10, &beta_slow,    sizeof(float));
-    memcpy(params + 11, &sections,     sizeof(int)*4);
+    if (mrope_used) {
+        memcpy(params + 11, sections,  sizeof(int32_t) * GGML_MROPE_SECTIONS);
+    } else {
+        memset(params + 11, 0,         sizeof(int32_t) * GGML_MROPE_SECTIONS);
+    }
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op     = GGML_OP_ROPE;
@@ -3934,7 +3946,7 @@ struct ggml_tensor * ggml_rope(
         int                   n_dims,
         int                   mode) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false
+        ctx, a, b, NULL, n_dims, NULL, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false
     );
 }
 
@@ -3944,7 +3956,7 @@ struct ggml_tensor * ggml_rope_multi(
         struct ggml_tensor  * b,
         struct ggml_tensor  * c,
         int                   n_dims,
-        int                   sections[4],
+        int                   sections[GGML_MROPE_SECTIONS],
         int                   mode,
         int                   n_ctx_orig,
         float                 freq_base,
@@ -3953,36 +3965,31 @@ struct ggml_tensor * ggml_rope_multi(
         float                 attn_factor,
         float                 beta_fast,
         float                 beta_slow) {
-    // Multimodal Rotary Position Embedding
-    GGML_ASSERT((mode & 1) == 0 && "mode & 1 == 1 is no longer supported");
-
-    GGML_ASSERT(ggml_is_vector(b));
-    GGML_ASSERT(b->type == GGML_TYPE_I32);
-    GGML_ASSERT(a->ne[2] * 4 == b->ne[0]); // mrope expecting 4 position ids per token
-
-    if (c) {
-        GGML_ASSERT(c->type == GGML_TYPE_F32);
-        GGML_ASSERT(c->ne[0] >= n_dims / 2);
-    }
-
-    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
-
-    int32_t params[11 + 4] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
-    memcpy(params +  5, &freq_base,    sizeof(float));
-    memcpy(params +  6, &freq_scale,   sizeof(float));
-    memcpy(params +  7, &ext_factor,   sizeof(float));
-    memcpy(params +  8, &attn_factor,  sizeof(float));
-    memcpy(params +  9, &beta_fast,    sizeof(float));
-    memcpy(params + 10, &beta_slow,    sizeof(float));
-    memcpy(&params[11], sections,      sizeof(int)*4);
-    ggml_set_op_params(result, params, sizeof(params));
-
-    result->op   = GGML_OP_ROPE;
-    result->src[0] = a;
-    result->src[1] = b;
-    result->src[2] = c;
+    return ggml_rope_impl(
+        ctx, a, b, c, n_dims, sections, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, false
+    );
+}
 
-    return result;
+struct ggml_tensor * ggml_rope_multi_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c,
+        int                   n_dims,
+        int                   sections[GGML_MROPE_SECTIONS],
+        int                   mode,
+        int                   n_ctx_orig,
+        float                 freq_base,
+        float                 freq_scale,
+        float                 ext_factor,
+        float                 attn_factor,
+        float                 beta_fast,
+        float                 beta_slow) {
+    return ggml_rope_impl(
+        ctx, a, b, c, n_dims, sections, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, true
+    );
 }
 
 struct ggml_tensor * ggml_rope_inplace(
@@ -3992,7 +3999,7 @@ struct ggml_tensor * ggml_rope_inplace(
         int                   n_dims,
         int                   mode) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true
+        ctx, a, b, NULL, n_dims, NULL, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true
     );
 }
 
@@ -4011,7 +4018,7 @@ struct ggml_tensor * ggml_rope_ext(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, c, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, false
     );
 }
@@ -4031,7 +4038,7 @@ struct ggml_tensor * ggml_rope_ext_inplace(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, c, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, true
     );
 }
@@ -4050,7 +4057,7 @@ struct ggml_tensor * ggml_rope_custom(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, NULL, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, false
     );
 }
@@ -4069,7 +4076,7 @@ struct ggml_tensor * ggml_rope_custom_inplace(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, NULL, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, true
     );
 }
@@ -4267,14 +4274,13 @@ struct ggml_tensor * ggml_conv_1d_dw(
         int                   s0,
         int                   p0,
         int                   d0) {
-    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], 1, a->ne[1], a->ne[2]);
     struct ggml_tensor * new_b = ggml_reshape_4d(ctx, b, b->ne[0], 1, b->ne[1], b->ne[2]);
 
-    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, new_b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16);
+    struct ggml_tensor * im2col = ggml_im2col(ctx, a, new_b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16);
 
     struct ggml_tensor * result = ggml_mul_mat(ctx, im2col, a);
 
-    result = ggml_reshape_3d(ctx, result, b->ne[0], b->ne[1], 1);
+    result = ggml_reshape_3d(ctx, result, result->ne[0], result->ne[2], 1);
 
     return result;
 }
@@ -4476,6 +4482,56 @@ struct ggml_tensor * ggml_conv_2d_direct(
     return result;
 }
 
+// ggml_conv_3d
+
+struct ggml_tensor * ggml_conv_3d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   s2,
+        int                   p0,
+        int                   p1,
+        int                   p2,
+        int                   d0,
+        int                   d1,
+        int                   d2,
+        int                   c,
+        int                   n,
+        int                   oc) {
+
+    GGML_ASSERT(a->ne[3] == (int64_t) c * oc);
+    GGML_ASSERT(b->ne[3] == (int64_t) c * n);
+
+    int64_t ne[4];
+    ne[0] = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+    ne[1] = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
+    ne[2] = ggml_calc_conv_output_size(b->ne[2], a->ne[2], s2, p2, d2);
+    ne[3] = (int64_t) oc * n;
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+
+    ggml_set_op_params_i32(result, 0,  s0);
+    ggml_set_op_params_i32(result, 1,  s1);
+    ggml_set_op_params_i32(result, 2,  s2);
+    ggml_set_op_params_i32(result, 3,  p0);
+    ggml_set_op_params_i32(result, 4,  p1);
+    ggml_set_op_params_i32(result, 5,  p2);
+    ggml_set_op_params_i32(result, 6,  d0);
+    ggml_set_op_params_i32(result, 7,  d1);
+    ggml_set_op_params_i32(result, 8,  d2);
+    ggml_set_op_params_i32(result, 9,  c);
+    ggml_set_op_params_i32(result, 10, n);
+    ggml_set_op_params_i32(result, 11, oc);
+
+    result->op = GGML_OP_CONV_3D;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
 // ggml_conv_transpose_2d_p0
 
 static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {
@@ -5602,6 +5658,28 @@ struct ggml_tensor * ggml_opt_step_adamw(
     return result;
 }
 
+// opt_step_sgd
+
+struct ggml_tensor * ggml_opt_step_sgd(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * grad,
+        struct ggml_tensor  * params) {
+    GGML_ASSERT(a->flags & GGML_TENSOR_FLAG_PARAM);
+    GGML_ASSERT(ggml_are_same_shape(a, grad));
+    GGML_ASSERT(params->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_nelements(params) == 2);
+
+    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
+
+    result->op     = GGML_OP_OPT_STEP_SGD;
+    result->src[0] = a;
+    result->src[1] = grad;
+    result->src[2] = params;
+
+    return result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 struct ggml_hash_set ggml_hash_set_new(size_t size) {
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 911eea504a19e..d03a02c7bf921 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -385,6 +385,7 @@ class MODEL_ARCH(IntEnum):
     DREAM            = auto()
     SMALLTHINKER     = auto()
     LLADA            = auto()
+    SEED_OSS         = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -717,6 +718,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.DREAM:            "dream",
     MODEL_ARCH.SMALLTHINKER:     "smallthinker",
     MODEL_ARCH.LLADA:            "llada",
+    MODEL_ARCH.SEED_OSS:         "seed_oss",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1973,6 +1975,20 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.SEED_OSS: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+    ],
     MODEL_ARCH.OLMOE: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2590,6 +2606,7 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.ATTN_K,
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.OUTPUT,
     ],
     MODEL_ARCH.SMALLTHINKER: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -2832,6 +2849,7 @@ class VisionProjectorType:
     QWEN2A = "qwen2a" # audio
     QWEN25O = "qwen2.5o" # omni
     VOXTRAL = "voxtral"
+    LFM2 = "lfm2"
 
 
 # Items here are (block size, type size)
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index dc7c03b464c25..87edaa3232ccc 100644
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -1119,7 +1119,8 @@ class TensorNameMap:
             "model.vision_tower.embeddings.patch_embeddings.projection", # Intern-S1
             "vpm.embeddings.patch_embedding",
             "model.vision_model.embeddings.patch_embedding", # SmolVLM
-            "vision_tower.patch_conv", # pixtral
+            "vision_tower.patch_conv", # pixtral-hf
+            "vision_encoder.patch_conv", # pixtral
             "vision_model.patch_embedding.linear", # llama 4
             "visual.patch_embed.proj", # qwen2vl
         ),
@@ -1138,7 +1139,8 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.q_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM
             "vision_model.model.layers.{bid}.self_attn.q_proj", # llama4
-            "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral
+            "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.attention.wq", # pixtral
             "visual.blocks.{bid}.attn.q", # qwen2vl, generated
         ),
 
@@ -1153,7 +1155,8 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.k_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM
             "vision_model.model.layers.{bid}.self_attn.k_proj", # llama4
-            "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral
+            "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.attention.wk", # pixtral
             "visual.blocks.{bid}.attn.k", # qwen2vl, generated
         ),
 
@@ -1168,7 +1171,8 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.v_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM
             "vision_model.model.layers.{bid}.self_attn.v_proj", # llama4
-            "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral
+            "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.attention.wv", # pixtral
             "visual.blocks.{bid}.attn.v", # qwen2vl, generated
         ),
 
@@ -1178,7 +1182,8 @@ class TensorNameMap:
             "model.vision_tower.encoder.layer.{bid}.layernorm_before", # Intern-S1
             "vpm.encoder.layers.{bid}.layer_norm1",
             "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM
-            "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral
+            "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.attention_norm", # pixtral
             "vision_model.model.layers.{bid}.input_layernorm", # llama4
             "visual.blocks.{bid}.norm1", # qwen2vl
         ),
@@ -1190,7 +1195,8 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.out_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
             "vision_model.model.layers.{bid}.self_attn.o_proj", # llama4
-            "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral
+            "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.attention.wo", # pixtral
             "visual.blocks.{bid}.attn.proj", # qwen2vl
         ),
 
@@ -1201,7 +1207,8 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.layer_norm2",
             "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM
             "vision_model.model.layers.{bid}.post_attention_layernorm", # llama4
-            "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral
+            "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.ffn_norm", # pixtral
             "visual.blocks.{bid}.norm2", # qwen2vl
         ),
 
@@ -1210,14 +1217,16 @@ class TensorNameMap:
             "model.vision_tower.encoder.layer.{bid}.mlp.fc1", # Intern-S1
             "vpm.encoder.layers.{bid}.mlp.fc1",
             "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3
-            "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
+            "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.feed_forward.w3", # pixtral
             "vision_model.model.layers.{bid}.mlp.fc1", # llama4
             "visual.blocks.{bid}.mlp.fc1", # qwen2vl
             "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl
         ),
 
         MODEL_TENSOR.V_ENC_FFN_GATE: (
-            "vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral
+            "vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.feed_forward.w1", # pixtral
             "visual.blocks.{bid}.mlp.gate_proj", # qwen2.5vl
         ),
 
@@ -1226,7 +1235,8 @@ class TensorNameMap:
             "model.vision_tower.encoder.layer.{bid}.mlp.fc2", # Intern-S1
             "vpm.encoder.layers.{bid}.mlp.fc2",
             "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3
-            "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
+            "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral-hf
+            "vision_encoder.transformer.layers.{bid}.feed_forward.w2", # pixtral
             "vision_model.model.layers.{bid}.mlp.fc2", # llama4
             "visual.blocks.{bid}.mlp.fc2", # qwen2vl
             "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
@@ -1244,7 +1254,8 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_PRE_NORM: (
             "vision_tower.vision_model.pre_layrnorm",
-            "vision_tower.ln_pre", # pixtral
+            "vision_tower.ln_pre", # pixtral-hf
+            "vision_encoder.ln_pre", # pixtral
             "vision_model.layernorm_pre", # llama4
         ),
 
@@ -1261,6 +1272,8 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MM_INP_NORM: (
             "multi_modal_projector.norm",
+            "multi_modal_projector.layer_norm",
+            "pre_mm_projector_norm",
         ),
 
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
@@ -1316,7 +1329,8 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.V_MM_PATCH_MERGER: (
-            "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1
+            "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1 - hf
+            "patch_merger.merging_layer", # mistral
         ),
 
         # audio (mtmd)
diff --git a/gguf-py/gguf/utility.py b/gguf-py/gguf/utility.py
index 00adcbc937398..769ccb02f0d91 100644
--- a/gguf-py/gguf/utility.py
+++ b/gguf-py/gguf/utility.py
@@ -145,7 +145,11 @@ def get_list_tensors_hf_model(cls, model_id: str) -> dict[str, RemoteTensor]:
                     tensors[key] = val
             return tensors
 
-        raise ValueError(f"Model {model_id} does not have any safetensor files")
+        raise ValueError(
+            f"No safetensor file has been found for model {model_id}."
+            "If the repo has safetensor files, make sure the model is public or you have a "
+            "valid Hugging Face token set in the environment variable HF_TOKEN."
+        )
 
     @classmethod
     def get_list_tensors(cls, url: str) -> dict[str, RemoteTensor]:
diff --git a/include/llama.h b/include/llama.h
index 545e957e5f52b..c5622cc16b4c2 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -64,8 +64,6 @@ extern "C" {
 
     typedef struct llama_memory_i * llama_memory_t;
 
-    struct llama_kv_cache; // DEPRECATED (use llama_memory instead)
-
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
     typedef int32_t llama_seq_id;
@@ -314,7 +312,7 @@ extern "C" {
         float    yarn_beta_fast;   // YaRN low correction dim
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
-        float    defrag_thold;     // defragment the KV cache if holes/size > thold, <= 0 disabled (default)
+        float    defrag_thold;     // [DEPRECATED] defragment the KV cache if holes/size > thold, <= 0 disabled (default)
 
         ggml_backend_sched_eval_callback cb_eval;
         void * cb_eval_user_data;
@@ -469,8 +467,6 @@ extern "C" {
     LLAMA_API           llama_memory_t   llama_get_memory  (const struct llama_context * ctx);
     LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type
 
-    DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead");
-
     LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
     LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
 
@@ -667,111 +663,6 @@ extern "C" {
     // Check if the memory supports shifting
     LLAMA_API bool llama_memory_can_shift(llama_memory_t mem);
 
-    //
-    // KV cache for self-attention (TODO: deprecate in favor of llama_memory)
-    //
-
-    // Returns the number of tokens in the KV cache (slow, use only for debug)
-    // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
-    DEPRECATED(LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx),
-               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
-
-    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    DEPRECATED(LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx),
-               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
-
-    // Clear the KV cache - both cell info is erased and KV data is zeroed
-    DEPRECATED(LLAMA_API void llama_kv_self_clear(
-                struct llama_context * ctx),
-            "Use llama_memory_clear() instead");
-
-    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
-    // seq_id < 0 : match any sequence
-    // p0 < 0     : [0,  p1]
-    // p1 < 0     : [p0, inf)
-    DEPRECATED(LLAMA_API bool llama_kv_self_seq_rm(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1),
-            "Use llama_memory_seq_rm() instead");
-
-    // Copy all tokens that belong to the specified sequence to another sequence
-    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_cp(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id_src,
-                    llama_seq_id   seq_id_dst,
-                       llama_pos   p0,
-                       llama_pos   p1),
-            "Use llama_memory_seq_cp() instead");
-
-    // Removes all tokens that do not belong to the specified sequence
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_keep(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_keep() instead");
-
-    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_add(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                       llama_pos   delta),
-            "Use llama_memory_seq_add() instead");
-
-    // Integer division of the positions by factor of `d > 1`
-    // If the KV cache is RoPEd, the KV data is updated accordingly:
-    //   - lazily on next llama_decode()
-    // p0 < 0 : [0,  p1]
-    // p1 < 0 : [p0, inf)
-    DEPRECATED(LLAMA_API void llama_kv_self_seq_div(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id,
-                       llama_pos   p0,
-                       llama_pos   p1,
-                             int   d),
-            "Use llama_memory_seq_div() instead");
-
-    // Returns the smallest position present in the KV cache for the specified sequence
-    // This is typically non-zero only for SWA caches
-    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
-    // Return -1 if the sequence is empty
-    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_min(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_pos_min() instead");
-
-    // Returns the largest position present in the KV cache for the specified sequence
-    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
-    // Return -1 if the sequence is empty
-    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_max(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id),
-            "Use llama_memory_seq_pos_max() instead");
-
-    // Defragment the KV cache
-    // This will be applied:
-    //   - lazily on next llama_decode()
-    DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx),
-            "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");
-
-    // Check if the context supports KV cache shifting
-    DEPRECATED(LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx),
-            "use llama_memory_can_shift() instead");
-
-    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx),
-            "simply remove this call, updates are applied lazily on the next llama_decode()");
-
     //
     // State / sessions
     //
@@ -870,6 +761,29 @@ extern "C" {
                           size_t   n_token_capacity,
                           size_t * n_token_count_out);
 
+#define LLAMA_STATE_SEQ_FLAGS_SWA_ONLY 1
+
+    typedef uint32_t llama_state_seq_flags;
+
+    LLAMA_API size_t llama_state_seq_get_size_ext(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_get_data_ext(
+            struct llama_context * ctx,
+                         uint8_t * dst,
+                          size_t   size,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_set_data_ext(
+            struct llama_context * ctx,
+                   const uint8_t * src,
+                          size_t   size,
+                    llama_seq_id   dest_seq_id,
+           llama_state_seq_flags   flags);
+
     //
     // Decoding
     //
@@ -1437,6 +1351,8 @@ extern "C" {
 
         ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
         void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+
+        enum ggml_opt_optimizer_type optimizer_type;
     };
 
     LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);
diff --git a/models/templates/openai-gpt-oss-120b.jinja b/models/templates/openai-gpt-oss-120b.jinja
new file mode 100644
index 0000000000000..dc7bb11927d29
--- /dev/null
+++ b/models/templates/openai-gpt-oss-120b.jinja
@@ -0,0 +1,331 @@
+{#-
+  In addition to the normal inputs of `messages` and `tools`, this template also accepts the
+  following kwargs:
+  - "builtin_tools": A list, can contain "browser" and/or "python".
+  - "model_identity": A string that optionally describes the model identity.
+  - "reasoning_effort": A string that describes the reasoning effort, defaults to "medium".
+ #}
+
+{#- Tool Definition Rendering ============================================== #}
+{%- macro render_typescript_type(param_spec, required_params, is_nullable=false) -%}
+    {%- if param_spec.type == "array" -%}
+        {%- if param_spec['items'] -%}
+            {%- if param_spec['items']['type'] == "string" -%}
+                {{- "string[]" }}
+            {%- elif param_spec['items']['type'] == "number" -%}
+                {{- "number[]" }}
+            {%- elif param_spec['items']['type'] == "integer" -%}
+                {{- "number[]" }}
+            {%- elif param_spec['items']['type'] == "boolean" -%}
+                {{- "boolean[]" }}
+            {%- else -%}
+                {%- set inner_type = render_typescript_type(param_spec['items'], required_params) -%}
+                {%- if inner_type == "object | object" or inner_type|length > 50 -%}
+                    {{- "any[]" }}
+                {%- else -%}
+                    {{- inner_type + "[]" }}
+                {%- endif -%}
+            {%- endif -%}
+            {%- if param_spec.nullable -%}
+                {{- " | null" }}
+            {%- endif -%}
+        {%- else -%}
+            {{- "any[]" }}
+            {%- if param_spec.nullable -%}
+                {{- " | null" }}
+            {%- endif -%}
+        {%- endif -%}
+    {%- elif param_spec.type is defined and param_spec.type is iterable and param_spec.type is not string and param_spec.type is not mapping and param_spec.type[0] is defined -%}
+        {#- Handle array of types like ["object", "object"] from Union[dict, list] #}
+        {%- if param_spec.type | length > 1 -%}
+            {{- param_spec.type | join(" | ") }}
+        {%- else -%}
+            {{- param_spec.type[0] }}
+        {%- endif -%}
+    {%- elif param_spec.oneOf -%}
+        {#- Handle oneOf schemas - check for complex unions and fallback to any #}
+        {%- set has_object_variants = false -%}
+        {%- for variant in param_spec.oneOf -%}
+            {%- if variant.type == "object" -%}
+                {%- set has_object_variants = true -%}
+            {%- endif -%}
+        {%- endfor -%}
+        {%- if has_object_variants and param_spec.oneOf|length > 1 -%}
+            {{- "any" }}
+        {%- else -%}
+            {%- for variant in param_spec.oneOf -%}
+                {{- render_typescript_type(variant, required_params) -}}
+                {%- if variant.description %}
+                    {{- "// " + variant.description }}
+                {%- endif -%}
+                {%- if variant.default is defined %}
+                    {{ "// default: " + variant.default|tojson }}
+                {%- endif -%}
+                {%- if not loop.last %}
+                    {{- " | " }}
+                {% endif -%}
+            {%- endfor -%}
+        {%- endif -%}
+    {%- elif param_spec.type == "string" -%}
+        {%- if param_spec.enum -%}
+            {{- '"' + param_spec.enum|join('" | "') + '"' -}}
+        {%- else -%}
+            {{- "string" }}
+            {%- if param_spec.nullable %}
+                {{- " | null" }}
+            {%- endif -%}
+        {%- endif -%}
+    {%- elif param_spec.type == "number" -%}
+        {{- "number" }}
+    {%- elif param_spec.type == "integer" -%}
+        {{- "number" }}
+    {%- elif param_spec.type == "boolean" -%}
+        {{- "boolean" }}
+
+    {%- elif param_spec.type == "object" -%}
+        {%- if param_spec.properties -%}
+            {{- "{\n" }}
+            {%- for prop_name, prop_spec in param_spec.properties.items() -%}
+                {{- prop_name -}}
+                {%- if prop_name not in (param_spec.required or []) -%}
+                    {{- "?" }}
+                {%- endif -%}
+                {{- ": " }}
+                {{ render_typescript_type(prop_spec, param_spec.required or []) }}
+                {%- if not loop.last -%}
+                    {{-", " }}
+                {%- endif -%}
+            {%- endfor -%}
+            {{- "}" }}
+        {%- else -%}
+            {{- "object" }}
+        {%- endif -%}
+    {%- else -%}
+        {{- "any" }}
+    {%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_tool_namespace(namespace_name, tools) -%}
+    {{- "## " + namespace_name + "\n\n" }}
+    {{- "namespace " + namespace_name + " {\n\n" }}
+    {%- for tool in tools %}
+        {%- set tool = tool.function %}
+        {{- "// " + tool.description + "\n" }}
+        {{- "type "+ tool.name + " = " }}
+        {%- if tool.parameters and tool.parameters.properties %}
+            {{- "(_: {\n" }}
+            {%- for param_name, param_spec in tool.parameters.properties.items() %}
+                {%- if param_spec.description %}
+                    {{- "// " + param_spec.description + "\n" }}
+                {%- endif %}
+                {{- param_name }}
+                {%- if param_name not in (tool.parameters.required or []) -%}
+                    {{- "?" }}
+                {%- endif -%}
+                {{- ": " }}
+                {{- render_typescript_type(param_spec, tool.parameters.required or []) }}
+                {%- if param_spec.default is defined -%}
+                    {%- if param_spec.enum %}
+                        {{- ", // default: " + param_spec.default }}
+                    {%- elif param_spec.oneOf %}
+                        {{- "// default: " + param_spec.default }}
+                    {%- else %}
+                        {{- ", // default: " + param_spec.default|tojson }}
+                    {%- endif -%}
+                {%- endif -%}
+                {%- if not loop.last %}
+                    {{- ",\n" }}
+                {%- else %}
+                    {{- ",\n" }}
+                {%- endif -%}
+            {%- endfor %}
+            {{- "}) => any;\n\n" }}
+        {%- else -%}
+            {{- "() => any;\n\n" }}
+        {%- endif -%}
+    {%- endfor %}
+    {{- "} // namespace " + namespace_name }}
+{%- endmacro -%}
+
+{%- macro render_builtin_tools(browser_tool, python_tool) -%}
+    {%- if browser_tool %}
+        {{- "## browser\n\n" }}
+        {{- "// Tool for browsing.\n" }}
+        {{- "// The `cursor` appears in brackets before each browsing display: `[{cursor}]`.\n" }}
+        {{- "// Cite information from the tool using the following format:\n" }}
+        {{- "// `【{cursor}†L{line_start}(-L{line_end})?】`, for example: `【6†L9-L11】` or `【8†L3】`.\n" }}
+        {{- "// Do not quote more than 10 words directly from the tool output.\n" }}
+        {{- "// sources=web (default: web)\n" }}
+        {{- "namespace browser {\n\n" }}
+        {{- "// Searches for information related to `query` and displays `topn` results.\n" }}
+        {{- "type search = (_: {\n" }}
+        {{- "query: string,\n" }}
+        {{- "topn?: number, // default: 10\n" }}
+        {{- "source?: string,\n" }}
+        {{- "}) => any;\n\n" }}
+        {{- "// Opens the link `id` from the page indicated by `cursor` starting at line number `loc`, showing `num_lines` lines.\n" }}
+        {{- "// Valid link ids are displayed with the formatting: `【{id}†.*】`.\n" }}
+        {{- "// If `cursor` is not provided, the most recent page is implied.\n" }}
+        {{- "// If `id` is a string, it is treated as a fully qualified URL associated with `source`.\n" }}
+        {{- "// If `loc` is not provided, the viewport will be positioned at the beginning of the document or centered on the most relevant passage, if available.\n" }}
+        {{- "// Use this function without `id` to scroll to a new location of an opened page.\n" }}
+        {{- "type open = (_: {\n" }}
+        {{- "id?: number | string, // default: -1\n" }}
+        {{- "cursor?: number, // default: -1\n" }}
+        {{- "loc?: number, // default: -1\n" }}
+        {{- "num_lines?: number, // default: -1\n" }}
+        {{- "view_source?: boolean, // default: false\n" }}
+        {{- "source?: string,\n" }}
+        {{- "}) => any;\n\n" }}
+        {{- "// Finds exact matches of `pattern` in the current page, or the page given by `cursor`.\n" }}
+        {{- "type find = (_: {\n" }}
+        {{- "pattern: string,\n" }}
+        {{- "cursor?: number, // default: -1\n" }}
+        {{- "}) => any;\n\n" }}
+        {{- "} // namespace browser\n\n" }}
+    {%- endif -%}
+
+    {%- if python_tool %}
+        {{- "## python\n\n" }}
+        {{- "Use this tool to execute Python code in your chain of thought. The code will not be shown to the user. This tool should be used for internal reasoning, but not for code that is intended to be visible to the user (e.g. when creating plots, tables, or files).\n\n" }}
+        {{- "When you send a message containing Python code to python, it will be executed in a stateful Jupyter notebook environment. python will respond with the output of the execution or time out after 120.0 seconds. The drive at '/mnt/data' can be used to save and persist user files. Internet access for this session is UNKNOWN. Depends on the cluster.\n\n" }}
+    {%- endif -%}
+{%- endmacro -%}
+
+{#- System Message Construction ============================================ #}
+{%- macro build_system_message() -%}
+    {%- if model_identity is not defined %}
+        {%- set model_identity = "You are ChatGPT, a large language model trained by OpenAI." %}
+    {%- endif %}
+    {{- model_identity + "\n" }}
+    {{- "Knowledge cutoff: 2024-06\n" }}
+    {{- "Current date: " + strftime_now("%Y-%m-%d") + "\n\n" }}
+    {%- if reasoning_effort is not defined %}
+        {%- set reasoning_effort = "medium" %}
+    {%- endif %}
+    {{- "Reasoning: " + reasoning_effort + "\n\n" }}
+    {%- if builtin_tools %}
+        {{- "# Tools\n\n" }}
+        {%- set available_builtin_tools = namespace(browser=false, python=false) %}
+        {%- for tool in builtin_tools %}
+            {%- if tool == "browser" %}
+                {%- set available_builtin_tools.browser = true %}
+            {%- elif tool == "python" %}
+                {%- set available_builtin_tools.python = true %}
+            {%- endif %}
+        {%- endfor %}
+        {{- render_builtin_tools(available_builtin_tools.browser, available_builtin_tools.python) }}
+    {%- endif -%}
+    {{- "# Valid channels: analysis, commentary, final. Channel must be included for every message." }}
+    {%- if tools -%}
+        {{- "\nCalls to these tools must go to the commentary channel: 'functions'." }}
+    {%- endif -%}
+{%- endmacro -%}
+
+{#- Main Template Logic ================================================= #}
+{#- Set defaults #}
+
+{#- Render system message #}
+{{- "<|start|>system<|message|>" }}
+{{- build_system_message() }}
+{{- "<|end|>" }}
+
+{#- Extract developer message #}
+{%- if messages[0].role == "developer" or messages[0].role == "system" %}
+    {%- set developer_message = messages[0].content %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set developer_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+
+{#- Render developer message #}
+{%- if developer_message or tools %}
+    {{- "<|start|>developer<|message|>" }}
+    {%- if developer_message %}
+        {{- "# Instructions\n\n" }}
+        {{- developer_message }}
+        {{- "\n\n" }}
+    {%- endif %}
+    {%- if tools -%}
+        {{- "# Tools\n\n" }}
+        {{- render_tool_namespace("functions", tools) }}
+    {%- endif -%}
+    {{- "<|end|>" }}
+{%- endif %}
+
+{#- Render messages #}
+{%- set last_tool_call = namespace(name=none) %}
+{%- for message in loop_messages -%}
+    {#- At this point only assistant/user/tool messages should remain #}
+    {%- if message.role == 'assistant' -%}
+        {#- Checks to ensure the messages are being passed in the format we expect #}
+        {%- if "content" in message %}
+            {%- if "<|channel|>analysis<|message|>" in message.content or "<|channel|>final<|message|>" in message.content %}
+                {{- raise_exception("You have passed a message containing <|channel|> tags in the content field. Instead of doing this, you should pass analysis messages (the string between '<|message|>' and '<|end|>') in the 'thinking' field, and final messages (the string between '<|message|>' and '<|end|>') in the 'content' field.") }}
+            {%- endif %}
+        {%- endif %}
+        {%- if "thinking" in message %}
+            {%- if "<|channel|>analysis<|message|>" in message.thinking or "<|channel|>final<|message|>" in message.thinking %}
+                {{- raise_exception("You have passed a message containing <|channel|> tags in the thinking field. Instead of doing this, you should pass analysis messages (the string between '<|message|>' and '<|end|>') in the 'thinking' field, and final messages (the string between '<|message|>' and '<|end|>') in the 'content' field.") }}
+            {%- endif %}
+        {%- endif %}
+        {%- if "tool_calls" in message %}
+            {#- We need very careful handling here - we want to drop the tool call analysis message if the model #}
+            {#- has output a later <|final|> message, but otherwise we want to retain it. This is the only case #}
+            {#- when we render CoT/analysis messages in inference. #}
+            {%- set future_final_message = namespace(found=false) %}
+            {%- for future_message in loop_messages[loop.index:] %}
+                {%- if future_message.role == 'assistant' and "tool_calls" not in future_message %}
+                    {%- set future_final_message.found = true %}
+                {%- endif %}
+            {%- endfor %}
+            {#- We assume max 1 tool call per message, and so we infer the tool call name #}
+            {#- in "tool" messages from the most recent assistant tool call name #}
+            {%- set tool_call = message.tool_calls[0] %}
+            {%- if tool_call.function %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {%- if message.content and message.thinking %}
+                {{- raise_exception("Cannot pass both content and thinking in an assistant message with tool calls! Put the analysis message in one or the other, but not both.") }}
+            {%- elif message.content and not future_final_message.found %}
+                {{- "<|start|>assistant<|channel|>analysis<|message|>" + message.content + "<|end|>" }}
+            {%- elif message.thinking and not future_final_message.found %}
+                {{- "<|start|>assistant<|channel|>analysis<|message|>" + message.thinking + "<|end|>" }}
+            {%- endif %}
+            {{- "<|start|>assistant to=" }}
+            {{- "functions." + tool_call.name + "<|channel|>commentary " }}
+            {{- (tool_call.content_type if tool_call.content_type is defined else "json") + "<|message|>" }}
+            {{- tool_call.arguments|tojson }}
+            {{- "<|call|>" }}
+            {%- set last_tool_call.name = tool_call.name %}
+        {%- elif loop.last and not add_generation_prompt %}
+            {#- Only render the CoT if the final turn is an assistant turn and add_generation_prompt is false #}
+            {#- This is a situation that should only occur in training, never in inference. #}
+            {%- if "thinking" in message %}
+                {{- "<|start|>assistant<|channel|>analysis<|message|>" + message.thinking + "<|end|>" }}
+            {%- endif %}
+            {#- <|return|> indicates the end of generation, but <|end|> does not #}
+            {#- <|return|> should never be an input to the model, but we include it as the final token #}
+            {#- when training, so the model learns to emit it. #}
+            {{- "<|start|>assistant<|channel|>final<|message|>" + message.content + "<|return|>" }}
+        {%- else %}
+            {#- CoT is dropped during all previous turns, so we never render it for inference #}
+            {{- "<|start|>assistant<|channel|>final<|message|>" + message.content + "<|end|>" }}
+            {%- set last_tool_call.name = none %}
+        {%- endif %}
+    {%- elif message.role == 'tool' -%}
+        {%- if last_tool_call.name is none %}
+            {{- raise_exception("Message has tool role, but there was no previous assistant message with a tool call!") }}
+        {%- endif %}
+        {{- "<|start|>functions." + last_tool_call.name }}
+        {{- " to=assistant<|channel|>commentary<|message|>" + message.content|tojson + "<|end|>" }}
+    {%- elif message.role == 'user' -%}
+        {{- "<|start|>user<|message|>" + message.content + "<|end|>" }}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- Generation prompt #}
+{%- if add_generation_prompt -%}
+<|start|>assistant
+{%- endif -%}
\ No newline at end of file
diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py
index 8366f89a08076..2cfca66e0938c 100755
--- a/scripts/compare-llama-bench.py
+++ b/scripts/compare-llama-bench.py
@@ -28,7 +28,6 @@
     "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
     "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
     "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
-    "defrag_thold",
     "use_mmap",     "embeddings",   "no_op_offload",  "n_prompt",   "n_gen",        "n_depth",
     "test_time",    "avg_ns",       "stddev_ns",      "avg_ts",     "stddev_ts",
 ]
@@ -38,7 +37,6 @@
     "TEXT",    "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER",
     "TEXT",    "INTEGER", "INTEGER", "TEXT",    "TEXT",    "INTEGER",
     "TEXT",    "INTEGER", "INTEGER", "INTEGER", "TEXT",    "TEXT",
-    "REAL",
     "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER",
     "TEXT",    "INTEGER", "INTEGER", "REAL",    "REAL",
 ]
diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh
index 29d30e0a188a1..826c560cd883a 100755
--- a/scripts/sync-ggml-am.sh
+++ b/scripts/sync-ggml-am.sh
@@ -74,21 +74,7 @@ while read c; do
         cmake/common.cmake \
         cmake/ggml-config.cmake.in \
         src/ggml-cpu/cmake/FindSIMD.cmake \
-        src/ggml*.h \
-        src/ggml*.c \
-        src/ggml*.cpp \
-        src/gguf*.cpp \
-        src/ggml-blas/* \
-        src/ggml-cann/* \
-        src/ggml-cpu/* \
-        src/ggml-cuda/* \
-        src/ggml-hip/* \
-        src/ggml-metal/* \
-        src/ggml-musa/* \
-        src/ggml-opencl/* \
-        src/ggml-rpc/* \
-        src/ggml-sycl/* \
-        src/ggml-vulkan/* \
+        src/ggml* \
         include/ggml*.h \
         include/gguf*.h \
         tests/test-opt.cpp \
@@ -131,21 +117,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # cmake/ggml-config.cmake.in        -> ggml/cmake/ggml-config.cmake.in
     # src/ggml-cpu/cmake/FindSIMD.cmake -> ggml/src/ggml-cpu/cmake/FindSIMD.cmake
     #
-    # src/ggml*.c          -> ggml/src/ggml*.c
-    # src/ggml*.cpp        -> ggml/src/ggml*.cpp
-    # src/ggml*.h          -> ggml/src/ggml*.h
-    # src/gguf*.cpp        -> ggml/src/gguf*.cpp
-    # src/ggml-blas/*      -> ggml/src/ggml-blas/*
-    # src/ggml-cann/*      -> ggml/src/ggml-cann/*
-    # src/ggml-cpu/*       -> ggml/src/ggml-cpu/*
-    # src/ggml-cuda/*      -> ggml/src/ggml-cuda/*
-    # src/ggml-hip/*       -> ggml/src/ggml-hip/*
-    # src/ggml-metal/*     -> ggml/src/ggml-metal/*
-    # src/ggml-musa/*      -> ggml/src/ggml-musa/*
-    # src/ggml-opencl/*    -> ggml/src/ggml-opencl/*
-    # src/ggml-rpc/*       -> ggml/src/ggml-rpc/*
-    # src/ggml-sycl/*      -> ggml/src/ggml-sycl/*
-    # src/ggml-vulkan/*    -> ggml/src/ggml-vulkan/*
+    # src/ggml* -> ggml/src/ggml*
     #
     # include/ggml*.h -> ggml/include/ggml*.h
     # include/gguf*.h -> ggml/include/gguf*.h
@@ -163,20 +135,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]| [ab]\/)cmake\/common.cmake/\1ggml\/cmake\/common.cmake/g' \
         -e 's/([[:space:]]| [ab]\/)cmake\/ggml-config.cmake.in/\1ggml\/cmake\/ggml-config.cmake.in/g' \
         -e 's/([[:space:]]| [ab]\/)src\/ggml-cpu\/cmake\/FindSIMD.cmake/\1ggml\/src\/ggml-cpu\/cmake\/FindSIMD.cmake/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.c/\1ggml\/src\/ggml\2.c/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.cpp/\1ggml\/src\/ggml\2.cpp/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.h/\1ggml\/src\/ggml\2.h/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/gguf(.*)\.cpp/\1ggml\/src\/gguf\2.cpp/g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-blas\//\1ggml\/src\/ggml-blas\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-cpu\//\1ggml\/src\/ggml-cpu\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-hip\//\1ggml\/src\/ggml-hip\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-metal\//\1ggml\/src\/ggml-metal\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-opencl\//\1ggml\/src\/ggml-opencl\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-rpc\//\1ggml\/src\/ggml-rpc\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \
-        -e 's/([[:space:]]| [ab]\/)src\/ggml-vulkan\//\1ggml\/src\/ggml-vulkan\//g' \
+        -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)/\1ggml\/src\/ggml\2/g' \
         -e 's/([[:space:]]| [ab]\/)include\/ggml(.*)\.h/\1ggml\/include\/ggml\2.h/g' \
         -e 's/([[:space:]]| [ab]\/)include\/gguf(.*)\.h/\1ggml\/include\/gguf\2.h/g' \
         -e 's/([[:space:]]| [ab]\/)tests\/(.*)\.cpp/\1tests\/\2.cpp/g' \
diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index 20a5915bc2be4..22348782d1911 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-daf7906728036a82f20c69fcbd74b6f536c74d3f
+323951f1bdcdfbd5b5ff3a9a7c3770e63b1a560e
diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh
index 9b98329e09cb6..2da9b57891582 100755
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@@ -6,21 +6,7 @@ cp -rpv ../ggml/src/CMakeLists.txt   ./ggml/src/CMakeLists.txt
 cp -rpv ../ggml/cmake/* ./ggml/cmake/
 cp -rpv ../ggml/src/ggml-cpu/cmake/* ./ggml/src/ggml-cpu/cmake/
 
-cp -rpv ../ggml/src/ggml*.c        ./ggml/src/
-cp -rpv ../ggml/src/ggml*.cpp      ./ggml/src/
-cp -rpv ../ggml/src/ggml*.h        ./ggml/src/
-cp -rpv ../ggml/src/gguf*.cpp      ./ggml/src/
-cp -rpv ../ggml/src/ggml-blas/*    ./ggml/src/ggml-blas/
-cp -rpv ../ggml/src/ggml-cann/*    ./ggml/src/ggml-cann/
-cp -rpv ../ggml/src/ggml-cpu/*     ./ggml/src/ggml-cpu/
-cp -rpv ../ggml/src/ggml-cuda/*    ./ggml/src/ggml-cuda/
-cp -rpv ../ggml/src/ggml-hip/*     ./ggml/src/ggml-hip/
-cp -rpv ../ggml/src/ggml-metal/*   ./ggml/src/ggml-metal/
-cp -rpv ../ggml/src/ggml-musa/*    ./ggml/src/ggml-musa/
-cp -rpv ../ggml/src/ggml-opencl/*  ./ggml/src/ggml-opencl/
-cp -rpv ../ggml/src/ggml-rpc/*     ./ggml/src/ggml-rpc/
-cp -rpv ../ggml/src/ggml-sycl/*    ./ggml/src/ggml-sycl/
-cp -rpv ../ggml/src/ggml-vulkan/*  ./ggml/src/ggml-vulkan/
+cp -rpv ../ggml/src/ggml* ./ggml/src/
 
 cp -rpv ../ggml/include/ggml*.h ./ggml/include/
 cp -rpv ../ggml/include/gguf*.h ./ggml/include/
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 8f9cd652447ab..18cfc76564d36 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -20,8 +20,8 @@ add_library(llama
             llama-hparams.cpp
             llama-impl.cpp
             llama-io.cpp
-            llama-kv-cache-unified.cpp
-            llama-kv-cache-unified-iswa.cpp
+            llama-kv-cache.cpp
+            llama-kv-cache-iswa.cpp
             llama-memory.cpp
             llama-memory-hybrid.cpp
             llama-memory-recurrent.cpp
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 18dcc6ddfe567..0ca0a4c22f814 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -93,6 +93,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DREAM,            "dream"            },
     { LLM_ARCH_SMALLTHINKER,     "smallthinker"     },
     { LLM_ARCH_LLADA,            "llada"            },
+    { LLM_ARCH_SEED_OSS,         "seed_oss"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -2010,6 +2011,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
             { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
             { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT,            "output" },
         }
     },
     {
@@ -2067,6 +2069,23 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_SEED_OSS,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 7af587e7951bc..7008c2514c5d4 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -97,6 +97,7 @@ enum llm_arch {
     LLM_ARCH_DREAM,
     LLM_ARCH_SMALLTHINKER,
     LLM_ARCH_LLADA,
+    LLM_ARCH_SEED_OSS,
     LLM_ARCH_UNKNOWN,
 };
 
diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp
index 8698d89acecb2..55d89eca0ad94 100644
--- a/src/llama-batch.cpp
+++ b/src/llama-batch.cpp
@@ -477,7 +477,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
 
 llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
     if (sequential && has_cpl) {
-        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag)\n", __func__);
 
         return {};
     }
diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp
index 0a96a9a579e26..9d8e57eac1f69 100644
--- a/src/llama-chat.cpp
+++ b/src/llama-chat.cpp
@@ -16,10 +16,10 @@
 static std::string trim(const std::string & str) {
     size_t start = 0;
     size_t end = str.size();
-    while (start < end && isspace(str[start])) {
+    while (start < end && isspace(static_cast<unsigned char>(str[start]))) {
         start += 1;
     }
-    while (end > start && isspace(str[end - 1])) {
+    while (end > start && isspace(static_cast<unsigned char>(str[end - 1]))) {
         end -= 1;
     }
     return str.substr(start, end - start);
@@ -69,6 +69,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "gpt-oss",           LLM_CHAT_TEMPLATE_OPENAI_MOE        },
     { "hunyuan-dense",     LLM_CHAT_TEMPLATE_HUNYUAN_DENSE     },
     { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
+    { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -201,6 +202,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_HUNYUAN_DENSE;
     } else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) {
         return LLM_CHAT_TEMPLATE_KIMI_K2;
+    } else if (tmpl_contains("<seed:bos>")) {
+        return LLM_CHAT_TEMPLATE_SEED_OSS;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -752,6 +755,14 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|im_assistant|>assistant<|im_middle|>";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SEED_OSS) {
+        for (auto message: chat) {
+            std::string role(message->role);
+            ss << "<seed:bos>" << role << "\n" << (role == "assistant" ? trim(message->content) : message->content) << "<seed:eos>";
+        }
+        if (add_ass) {
+            ss << "<seed:bos>assistant\n";
+        }
     } else {
         // template not supported
         return -1;
diff --git a/src/llama-chat.h b/src/llama-chat.h
index 35a943856fa52..21d53ed08b4c3 100644
--- a/src/llama-chat.h
+++ b/src/llama-chat.h
@@ -49,6 +49,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_OPENAI_MOE,
     LLM_CHAT_TEMPLATE_HUNYUAN_DENSE,
     LLM_CHAT_TEMPLATE_KIMI_K2,
+    LLM_CHAT_TEMPLATE_SEED_OSS,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 26a5cf9c3f8db..18cf25079d283 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -39,7 +39,6 @@ llama_context::llama_context(
     cparams.yarn_attn_factor = params.yarn_attn_factor;
     cparams.yarn_beta_fast   = params.yarn_beta_fast;
     cparams.yarn_beta_slow   = params.yarn_beta_slow;
-    cparams.defrag_thold     = params.defrag_thold;
     cparams.embeddings       = params.embeddings;
     cparams.offload_kqv      = params.offload_kqv;
     cparams.flash_attn       = params.flash_attn;
@@ -93,7 +92,7 @@ llama_context::llama_context(
     // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
     // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
     // ref: https://github.com/ggerganov/llama.cpp/pull/5021
-    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_context_kv_self
+    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_memory
     if (cparams.n_batch < GGML_KQ_MASK_PAD) {
         LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
         cparams.n_batch = GGML_KQ_MASK_PAD;
@@ -145,11 +144,6 @@ llama_context::llama_context(
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
-    if (!params.swa_full && cparams.n_seq_max > 1 && hparams.is_swa_any()) {
-        LLAMA_LOG_WARN("%s: requested n_seq_max (%u) > 1, but swa_full is not enabled -- performance may be degraded: %s\n",
-                __func__, cparams.n_seq_max, "https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573");
-    }
-
     if (!hparams.vocab_only) {
         // GPU backends
         for (auto * dev : model.devices) {
@@ -444,26 +438,12 @@ llama_memory_t llama_context::get_memory() const {
     return memory.get();
 }
 
-// deprecated
-void llama_context::kv_self_defrag_sched() {
-    if (!memory) {
-        return;
-    }
-
-    memory_force_optimize = true;
-}
-
-// deprecated
-bool llama_context::kv_self_update(bool optimize) {
+bool llama_context::memory_update(bool optimize) {
     if (!memory) {
         return false;
     }
 
     {
-        // TODO: remove in the future
-        optimize |= memory_force_optimize;
-        memory_force_optimize = false;
-
         const auto mctx = memory->init_update(this, optimize);
         switch (mctx->get_status()) {
             case LLAMA_MEMORY_STATUS_SUCCESS:
@@ -997,8 +977,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
 
     bool did_optimize = false;
 
-    // handle any pending defrags/shifts
-    kv_self_update(false);
+    // handle any pending shifts/copies
+    memory_update(false);
 
     llama_memory_context_ptr mctx;
 
@@ -1023,7 +1003,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     if (!did_optimize) {
                         did_optimize = true;
 
-                        if (kv_self_update(true)) {
+                        if (memory_update(true)) {
                             LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, balloc->get_n_tokens());
 
                             continue;
@@ -1657,30 +1637,30 @@ size_t llama_context::state_set_data(const uint8_t * src, size_t size) {
     }
 }
 
-size_t llama_context::state_seq_get_size(llama_seq_id seq_id) {
+size_t llama_context::state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags) {
     llama_io_write_dummy io;
     try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
         return 0;
     }
 }
 
-size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size) {
+size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size, llama_state_seq_flags flags) {
     llama_io_write_buffer io(dst, size);
     try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
         return 0;
     }
 }
 
-size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size) {
+size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags) {
     llama_io_read_buffer io(src, size);
     try {
-        return state_seq_read_data(io, seq_id);
+        return state_seq_read_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
         return 0;
@@ -1778,7 +1758,7 @@ size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * file
     {
         const size_t state_size = file.size() - file.tell();
         llama_io_read_file io(&file);
-        const size_t nread = state_seq_read_data(io, seq_id);
+        const size_t nread = state_seq_read_data(io, seq_id, 0);
         if (!nread) {
             LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
             return 0;
@@ -1802,7 +1782,7 @@ size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * file
 
     // save the context state using stream saving
     llama_io_write_file io(&file);
-    state_seq_write_data(io, seq_id);
+    state_seq_write_data(io, seq_id, 0);
 
     const size_t res = file.tell();
     GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + io.n_bytes());
@@ -1971,21 +1951,21 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
     return io.n_bytes();
 }
 
-size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        memory->state_write(io, seq_id);
+        memory->state_write(io, seq_id, flags);
     }
 
     return io.n_bytes();
 }
 
-size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        memory->state_read(io, seq_id);
+        memory->state_read(io, seq_id, flags);
     }
 
     return io.n_bytes();
@@ -2048,7 +2028,7 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
     opt_params.opt_period      = n_batch / n_ubatch;
     opt_params.get_opt_pars    = lopt_params.get_opt_pars;
     opt_params.get_opt_pars_ud = lopt_params.get_opt_pars_ud;
-
+    opt_params.optimizer       = lopt_params.optimizer_type;
     opt_ctx = ggml_opt_init(opt_params);
 
     llama_opt_param_filter param_filter = lopt_params.param_filter;
@@ -2343,16 +2323,6 @@ const llama_model * llama_get_model(const llama_context * ctx) {
     return &ctx->get_model();
 }
 
-// deprecated
-llama_kv_cache * llama_get_kv_self(llama_context * ctx) {
-    return dynamic_cast<llama_kv_cache *>(ctx->get_memory());
-}
-
-// deprecated
-void llama_kv_self_update(llama_context * ctx) {
-    ctx->kv_self_update(false);
-}
-
 enum llama_pooling_type llama_pooling_type(const llama_context * ctx) {
     return ctx->pooling_type();
 }
@@ -2570,168 +2540,6 @@ bool llama_memory_can_shift(llama_memory_t mem) {
     return mem->get_can_shift();
 }
 
-//
-// kv cache
-//
-
-// deprecated
-int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
-    const auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return 0;
-    }
-
-    int32_t res = 0;
-
-    for (uint32_t s = 0; s < ctx->get_cparams().n_seq_max; s++) {
-        const llama_pos p0 = kv->seq_pos_min(s);
-        const llama_pos p1 = kv->seq_pos_max(s);
-
-        if (p0 >= 0) {
-            res += (p1 - p0) + 1;
-        }
-    }
-
-    return res;
-}
-
-// deprecated
-// note: this is the same as above - will be removed anyway, so it's ok
-int32_t llama_kv_self_used_cells(const llama_context * ctx) {
-    const auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return 0;
-    }
-
-    int32_t res = 0;
-
-    for (uint32_t s = 0; s < ctx->get_cparams().n_seq_max; s++) {
-        const llama_pos p0 = kv->seq_pos_min(s);
-        const llama_pos p1 = kv->seq_pos_max(s);
-
-        if (p0 >= 0) {
-            res += (p1 - p0) + 1;
-        }
-    }
-
-    return res;
-}
-
-// deprecated
-void llama_kv_self_clear(llama_context * ctx) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_clear(kv, true);
-}
-
-// deprecated
-bool llama_kv_self_seq_rm(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return true;
-    }
-
-    return llama_memory_seq_rm(kv, seq_id, p0, p1);
-}
-
-// deprecated
-void llama_kv_self_seq_cp(
-        llama_context * ctx,
-         llama_seq_id   seq_id_src,
-         llama_seq_id   seq_id_dst,
-            llama_pos   p0,
-            llama_pos   p1) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_cp(kv, seq_id_src, seq_id_dst, p0, p1);
-}
-
-// deprecated
-void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_keep(kv, seq_id);
-}
-
-// deprecated
-void llama_kv_self_seq_add(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1,
-            llama_pos   delta) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_add(kv, seq_id, p0, p1, delta);
-}
-
-// deprecated
-void llama_kv_self_seq_div(
-        llama_context * ctx,
-         llama_seq_id   seq_id,
-            llama_pos   p0,
-            llama_pos   p1,
-                  int   d) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return;
-    }
-
-    llama_memory_seq_div(kv, seq_id, p0, p1, d);
-}
-
-// deprecated
-llama_pos llama_kv_self_seq_pos_min(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return -1;
-    }
-
-    return llama_memory_seq_pos_min(kv, seq_id);
-}
-
-// deprecated
-llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return -1;
-    }
-
-    return llama_memory_seq_pos_max(kv, seq_id);
-}
-
-// deprecated
-void llama_kv_self_defrag(llama_context * ctx) {
-    // force defrag
-    ctx->kv_self_defrag_sched();
-}
-
-// deprecated
-bool llama_kv_self_can_shift(const llama_context * ctx) {
-    auto * kv = llama_get_memory(ctx);
-    if (!kv) {
-        return false;
-    }
-
-    return llama_memory_can_shift(kv);
-}
-
 // llama state API
 
 // deprecated
@@ -2801,19 +2609,31 @@ bool llama_state_save_file(llama_context * ctx, const char * path_session, const
 }
 
 size_t llama_state_seq_get_size(llama_context * ctx, llama_seq_id seq_id) {
-    return ctx->state_seq_get_size(seq_id);
+    return llama_state_seq_get_size_ext(ctx, seq_id, 0);
 }
 
 size_t llama_state_seq_get_data(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) {
+    return llama_state_seq_get_data_ext(ctx, dst, size, seq_id, 0);
+}
+
+size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) {
+    return llama_state_seq_set_data_ext(ctx, src, size, seq_id, 0);
+}
+
+size_t llama_state_seq_get_size_ext(llama_context * ctx, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    return ctx->state_seq_get_size(seq_id, flags);
+}
+
+size_t llama_state_seq_get_data_ext(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
     ctx->synchronize();
 
-    return ctx->state_seq_get_data(seq_id, dst, size);
+    return ctx->state_seq_get_data(seq_id, dst, size, flags);
 }
 
-size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) {
+size_t llama_state_seq_set_data_ext(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
     ctx->synchronize();
 
-    return ctx->state_seq_set_data(seq_id, src, size);
+    return ctx->state_seq_set_data(seq_id, src, size, flags);
 }
 
 size_t llama_state_seq_save_file(llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {
diff --git a/src/llama-context.h b/src/llama-context.h
index 25c143d56dfb2..3dd9205446483 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -46,10 +46,8 @@ struct llama_context {
 
     llama_memory_t get_memory() const;
 
-    // return true of the KV cache was updated
-    // TODO: remove
-    bool kv_self_update(bool optimize);
-    void kv_self_defrag_sched();
+    // return true if the memory was updated
+    bool memory_update(bool optimize);
 
     enum llama_pooling_type pooling_type() const;
 
@@ -111,9 +109,9 @@ struct llama_context {
     size_t state_get_data(      uint8_t * dst, size_t size);
     size_t state_set_data(const uint8_t * src, size_t size);
 
-    size_t state_seq_get_size(llama_seq_id seq_id);
-    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
-    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+    size_t state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size, llama_state_seq_flags flags);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags);
 
     bool state_load_file(
             const char * filepath,
@@ -152,6 +150,7 @@ struct llama_context {
 
     void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
 
+    // TODO: more flexible combinations of logical/physical batch size and context size
     void opt_epoch(
             ggml_opt_dataset_t      dataset,
             ggml_opt_result_t       result_train,
@@ -212,8 +211,8 @@ struct llama_context {
     size_t state_write_data(llama_io_write_i & io);
     size_t state_read_data (llama_io_read_i  & io);
 
-    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
-    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id, llama_state_seq_flags flags);
 
     //
     // members
@@ -229,9 +228,6 @@ struct llama_context {
 
     std::unique_ptr<llama_memory_i> memory;
 
-    // TODO: temporary, until the llama_kv_self_defrag() API is removed
-    bool memory_force_optimize = false;
-
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
     float * logits      = nullptr;
diff --git a/src/llama-cparams.h b/src/llama-cparams.h
index 38750affc500b..dbbaba9f6274c 100644
--- a/src/llama-cparams.h
+++ b/src/llama-cparams.h
@@ -24,7 +24,6 @@ struct llama_cparams {
     float yarn_attn_factor;
     float yarn_beta_fast;
     float yarn_beta_slow;
-    float defrag_thold;
 
     bool embeddings;
     bool causal_attn;
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 053c72d6dc8d1..6419d739bd8a2 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -4,8 +4,8 @@
 #include "llama-batch.h"
 #include "llama-cparams.h"
 
-#include "llama-kv-cache-unified.h"
-#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
 #include "llama-memory-recurrent.h"
 
@@ -277,7 +277,7 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                 for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
                     const llama_seq_id s0 = ubatch->seq_id[i0][0];
 
-                    // TODO: reimplement this like in llama_kv_cache_unified
+                    // TODO: reimplement this like in llama_kv_cache
                     if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) {
                         if (hparams.use_alibi) {
                             f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
@@ -294,15 +294,15 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {
     mctx->set_input_k_idxs(self_k_idxs, ubatch);
     mctx->set_input_v_idxs(self_v_idxs, ubatch);
 
     mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
 }
 
-bool llm_graph_input_attn_kv_unified::can_reuse(const llm_graph_params & params) {
-    const auto * mctx = static_cast<const llama_kv_cache_unified_context *>(params.mctx);
+bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);
 
     this->mctx = mctx;
 
@@ -319,7 +319,7 @@ bool llm_graph_input_attn_kv_unified::can_reuse(const llm_graph_params & params)
     return res;
 }
 
-void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
     mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
     mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
 
@@ -331,8 +331,8 @@ void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch
     mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
 }
 
-bool llm_graph_input_attn_kv_unified_iswa::can_reuse(const llm_graph_params & params) {
-    const auto * mctx = static_cast<const llama_kv_cache_unified_iswa_context *>(params.mctx);
+bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_iswa_context *>(params.mctx);
 
     this->mctx = mctx;
 
@@ -1186,7 +1186,7 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);
 
     auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, mctx_cur);
 
@@ -1223,8 +1223,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(
          ggml_tensor * v,
          ggml_tensor * kq_b,
          ggml_tensor * kq_mask,
-         ggml_tensor * v_mla,
          ggml_tensor * sinks,
+         ggml_tensor * v_mla,
              float     kq_scale) const {
     const bool v_trans = v->nb[1] > v->nb[2];
 
@@ -1360,6 +1360,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * sinks,
         ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
@@ -1381,7 +1382,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1399,17 +1400,17 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-static std::unique_ptr<llm_graph_input_attn_kv_unified> build_attn_inp_kv_unified_impl(
+static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
            ggml_context * ctx0,
      const llama_ubatch & ubatch,
     const llama_hparams & hparams,
     const llama_cparams & cparams,
-    const llama_kv_cache_unified_context * mctx_cur) {
+    const llama_kv_cache_context * mctx_cur) {
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, mctx_cur);
+    auto inp = std::make_unique<llm_graph_input_attn_kv>(hparams, cparams, mctx_cur);
 
     {
-        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");
 
         const auto n_kv     = mctx_cur->get_n_kv();
         const auto n_tokens = ubatch.n_tokens;
@@ -1427,22 +1428,23 @@ static std::unique_ptr<llm_graph_input_attn_kv_unified> build_attn_inp_kv_unifie
     return inp;
 }
 
-llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+llm_graph_input_attn_kv * llm_graph_context::build_attn_inp_kv() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);
 
-    auto inp = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
+    auto inp = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
 
-    return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
+    return (llm_graph_input_attn_kv *) res->add_input(std::move(inp));
 }
 
 ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_attn_kv_unified * inp,
+        llm_graph_input_attn_kv * inp,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * sinks,
         ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
@@ -1469,7 +1471,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1488,40 +1490,15 @@ ggml_tensor * llm_graph_context::build_attn(
 }
 
 ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_attn_kv_unified_iswa * inp,
-        ggml_tensor * wo,
-        ggml_tensor * wo_b,
-        ggml_tensor * q_cur,
-        ggml_tensor * k_cur,
-        ggml_tensor * v_cur,
-        ggml_tensor * kq_b,
-        ggml_tensor * v_mla,
-            float     kq_scale,
-            int       il) const {
-    return build_attn_with_sinks(
-            inp,
-            wo,
-            wo_b,
-            q_cur,
-            k_cur,
-            v_cur,
-            kq_b,
-            v_mla,
-            nullptr,
-            kq_scale,
-            il);
-}
-
-ggml_tensor * llm_graph_context::build_attn_with_sinks(
-        llm_graph_input_attn_kv_unified_iswa * inp,
+        llm_graph_input_attn_kv_iswa * inp,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
-        ggml_tensor * v_mla,
         ggml_tensor * sinks,
+        ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
@@ -1561,7 +1538,7 @@ ggml_tensor * llm_graph_context::build_attn_with_sinks(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, sinks, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1600,6 +1577,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * sinks,
         ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
@@ -1615,7 +1593,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, nullptr, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1636,10 +1614,10 @@ ggml_tensor * llm_graph_context::build_attn(
 // TODO: maybe separate the inner implementation into a separate function
 //       like with the non-sliding window equivalent
 //       once sliding-window hybrid caches are a thing.
-llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
+llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_iswa_context *>(mctx);
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
+    auto inp = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, mctx_cur);
 
     const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
@@ -1656,7 +1634,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
     }
 
     {
-        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
+        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache for non-SWA");
 
         const auto n_kv = mctx_cur->get_swa()->get_n_kv();
 
@@ -1669,7 +1647,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
     }
 
-    return (llm_graph_input_attn_kv_unified_iswa *) res->add_input(std::move(inp));
+    return (llm_graph_input_attn_kv_iswa *) res->add_input(std::move(inp));
 }
 
 ggml_tensor * llm_graph_context::build_rs(
@@ -1792,7 +1770,7 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx);
 
     auto inp_rs   = build_rs_inp_impl(ctx0, ubatch, mctx_cur->get_recr());
-    auto inp_attn = build_attn_inp_kv_unified_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
+    auto inp_attn = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
 
     auto inp = std::make_unique<llm_graph_input_mem_hybrid>(std::move(inp_attn), std::move(inp_rs), mctx_cur);
 
diff --git a/src/llama-graph.h b/src/llama-graph.h
index 6ff49de3a1ce8..e11d91d5293f0 100644
--- a/src/llama-graph.h
+++ b/src/llama-graph.h
@@ -19,8 +19,8 @@ struct llama_cparams;
 
 struct llama_memory_context_i;
 
-class llama_kv_cache_unified_context;
-class llama_kv_cache_unified_iswa_context;
+class llama_kv_cache_context;
+class llama_kv_cache_iswa_context;
 class llama_memory_recurrent_context;
 class llama_memory_hybrid_context;
 
@@ -152,7 +152,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
 public:
     llm_graph_input_pos_bucket_kv(
             const llama_hparams & hparams,
-            const llama_kv_cache_unified_context * mctx) : hparams(hparams), mctx(mctx) {}
+            const llama_kv_cache_context * mctx) : hparams(hparams), mctx(mctx) {}
     virtual ~llm_graph_input_pos_bucket_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -161,7 +161,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
 
     const llama_hparams hparams;
 
-    const llama_kv_cache_unified_context * mctx;
+    const llama_kv_cache_context * mctx;
 };
 
 class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -257,17 +257,17 @@ class llm_graph_input_attn_no_cache : public llm_graph_input_i {
     const llama_cparams cparams;
 };
 
-class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
+class llm_graph_input_attn_kv : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_kv_unified(
+    llm_graph_input_attn_kv(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_context * mctx) :
+            const llama_kv_cache_context * mctx) :
         hparams(hparams),
         cparams(cparams),
         mctx(mctx) {
     }
-    ~llm_graph_input_attn_kv_unified() = default;
+    ~llm_graph_input_attn_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
@@ -290,20 +290,20 @@ class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
     const llama_hparams hparams;
     const llama_cparams cparams;
 
-    const llama_kv_cache_unified_context * mctx;
+    const llama_kv_cache_context * mctx;
 };
 
-class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
+class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_kv_unified_iswa(
+    llm_graph_input_attn_kv_iswa(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_iswa_context * mctx) :
+            const llama_kv_cache_iswa_context * mctx) :
         hparams(hparams),
         cparams(cparams),
         mctx(mctx) {
     }
-    ~llm_graph_input_attn_kv_unified_iswa() = default;
+    ~llm_graph_input_attn_kv_iswa() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
@@ -330,7 +330,7 @@ class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
     const llama_hparams hparams;
     const llama_cparams cparams;
 
-    const llama_kv_cache_unified_iswa_context * mctx;
+    const llama_kv_cache_iswa_context * mctx;
 };
 
 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -351,7 +351,7 @@ class llm_graph_input_attn_cross : public llm_graph_input_i {
 class llm_graph_input_mem_hybrid : public llm_graph_input_i {
 public:
     llm_graph_input_mem_hybrid(
-            std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn,
+            std::unique_ptr<llm_graph_input_attn_kv> inp_attn,
             std::unique_ptr<llm_graph_input_rs>              inp_rs,
             const llama_memory_hybrid_context *              mctx) :
         inp_attn(std::move(inp_attn)),
@@ -361,11 +361,11 @@ class llm_graph_input_mem_hybrid : public llm_graph_input_i {
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn;
-    std::unique_ptr<llm_graph_input_rs>              inp_rs;
+    std::unique_ptr<llm_graph_input_attn_kv> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>      inp_rs;
 
-    llm_graph_input_attn_kv_unified * get_attn() const { return inp_attn.get(); }
-    llm_graph_input_rs              * get_recr() const { return inp_rs.get(); }
+    llm_graph_input_attn_kv * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs      * get_recr() const { return inp_rs.get(); }
 
     const llama_memory_hybrid_context * mctx;
 };
@@ -680,14 +680,14 @@ struct llm_graph_context {
     //
 
     ggml_tensor * build_attn_mha(
-             ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
-             ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
-             ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
-             ggml_tensor * kq_b,
-             ggml_tensor * kq_mask,
-             ggml_tensor * sinks,
-             ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                   float   kq_scale) const;
+            ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
+            ggml_tensor * kq_b,
+            ggml_tensor * kq_mask,
+            ggml_tensor * sinks,   // [n_head_q]
+            ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale) const;
 
     llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
 
@@ -699,50 +699,39 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
-    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
+    llm_graph_input_attn_kv * build_attn_inp_kv() const;
 
     ggml_tensor * build_attn(
-            llm_graph_input_attn_kv_unified * inp,
+            llm_graph_input_attn_kv * inp,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
-    llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
+    llm_graph_input_attn_kv_iswa * build_attn_inp_kv_iswa() const;
 
     // note: if k_cur or v_cur are not provided, they will not be stored in the memory
     ggml_tensor * build_attn(
-            llm_graph_input_attn_kv_unified_iswa * inp,
-            ggml_tensor * wo,
-            ggml_tensor * wo_b,
-            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
-            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
-            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
-            ggml_tensor * kq_b,
-            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                  float   kq_scale,
-                    int   il) const;
-
-    // TODO: temporary to keep the diff small. after the code is public will refactor to simplify this
-    ggml_tensor * build_attn_with_sinks(
-            llm_graph_input_attn_kv_unified_iswa * inp,
+            llm_graph_input_attn_kv_iswa * inp,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
             ggml_tensor * kq_b,
-            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
             ggml_tensor * sinks, // [n_head_q]
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
@@ -756,6 +745,7 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
@@ -765,7 +755,7 @@ struct llm_graph_context {
     //
 
     // TODO: move this implementation to llama_memory_recurrent.
-    //       this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
+    //       this is analogous to llama_kv_cache::cpy_k / cpy_v
     //       when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
     //         implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
     //         `llama_memory_recurrent`
diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index 7a06368dcda68..91636572da8b2 100644
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -153,3 +153,28 @@ bool llama_hparams::is_swa(uint32_t il) const {
 
     GGML_ABORT("fatal error");
 }
+
+bool llama_hparams::has_kv(uint32_t il) const {
+    if (n_layer_kv_from_start >= 0) {
+        if (il < (uint32_t) n_layer_kv_from_start) {
+            return true;
+        }
+
+        return false;
+    }
+
+    // by default, all layers have kv
+    return true;
+}
+
+uint32_t llama_hparams::n_layer_kv() const {
+    uint32_t res = 0;
+
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (has_kv(il)) {
+            res++;
+        }
+    }
+
+    return res;
+}
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index bd23122443271..60415f0c202a4 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -41,6 +41,7 @@ struct llama_hparams {
     uint32_t n_embd;
     uint32_t n_embd_features = 0;
     uint32_t n_layer;
+     int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
     uint32_t n_rot;
     uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
     uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
@@ -221,6 +222,11 @@ struct llama_hparams {
     uint32_t n_pos_per_embd() const;
 
     bool is_swa(uint32_t il) const;
+
+    bool has_kv(uint32_t il) const;
+
+    // number of layers for which has_kv() returns true
+    uint32_t n_layer_kv() const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
diff --git a/src/llama-kv-cache-unified-iswa.cpp b/src/llama-kv-cache-iswa.cpp
similarity index 57%
rename from src/llama-kv-cache-unified-iswa.cpp
rename to src/llama-kv-cache-iswa.cpp
index 01d27fb4db9b1..d7342914c6b7c 100644
--- a/src/llama-kv-cache-unified-iswa.cpp
+++ b/src/llama-kv-cache-iswa.cpp
@@ -1,4 +1,4 @@
-#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache-iswa.h"
 
 #include "llama-impl.h"
 #include "llama-batch.h"
@@ -8,10 +8,10 @@
 #include <cassert>
 
 //
-// llama_kv_cache_unified_iswa
+// llama_kv_cache_iswa
 //
 
-llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
+llama_kv_cache_iswa::llama_kv_cache_iswa(
         const llama_model & model,
                 ggml_type   type_k,
                 ggml_type   type_v,
@@ -22,9 +22,26 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
                  uint32_t   kv_size,
                  uint32_t   n_seq_max,
                  uint32_t   n_ubatch,
-                 uint32_t   n_pad) : hparams(model.hparams), unified(unified) {
-    llama_kv_cache_unified::layer_filter_cb filter_base = [&](int32_t il) { return !model.hparams.is_swa(il); };
-    llama_kv_cache_unified::layer_filter_cb filter_swa  = [&](int32_t il) { return  model.hparams.is_swa(il); };
+                 uint32_t   n_pad,
+    const layer_filter_cb & filter,
+    const  layer_reuse_cb & reuse) : hparams(model.hparams), unified(unified) {
+
+    // chain filters
+    const layer_filter_cb filter_base = [&](int32_t il) {
+        if (filter && !filter(il)) {
+            return false;
+        }
+
+        return !model.hparams.is_swa(il);
+    };
+
+    const layer_filter_cb filter_swa  = [&](int32_t il) {
+        if (filter && !filter(il)) {
+            return false;
+        }
+
+        return  model.hparams.is_swa(il);
+    };
 
     const uint32_t size_base = kv_size;
 
@@ -40,25 +57,25 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
 
     LLAMA_LOG_INFO("%s: creating non-SWA KV cache, size = %u cells\n", __func__, size_base);
 
-    kv_base = std::make_unique<llama_kv_cache_unified>(
-            model, std::move(filter_base), type_k, type_v,
+    kv_base = std::make_unique<llama_kv_cache>(
+            model, type_k, type_v,
             v_trans, offload, unified, size_base, n_seq_max, n_pad,
-            0, LLAMA_SWA_TYPE_NONE);
+            0, LLAMA_SWA_TYPE_NONE, filter_base, reuse);
 
     LLAMA_LOG_INFO("%s: creating     SWA KV cache, size = %u cells\n", __func__, size_swa);
 
-    kv_swa = std::make_unique<llama_kv_cache_unified>(
-            model, std::move(filter_swa), type_k, type_v,
+    kv_swa = std::make_unique<llama_kv_cache>(
+            model, type_k, type_v,
             v_trans, offload, unified, size_swa, n_seq_max, n_pad,
-            hparams.n_swa, hparams.swa_type);
+            hparams.n_swa, hparams.swa_type, filter_swa, reuse);
 }
 
-void llama_kv_cache_unified_iswa::clear(bool data) {
+void llama_kv_cache_iswa::clear(bool data) {
     kv_base->clear(data);
     kv_swa ->clear(data);
 }
 
-bool llama_kv_cache_unified_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+bool llama_kv_cache_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
     bool res = true;
 
     res = res & kv_base->seq_rm(seq_id, p0, p1);
@@ -67,36 +84,36 @@ bool llama_kv_cache_unified_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llam
     return res;
 }
 
-void llama_kv_cache_unified_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+void llama_kv_cache_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     kv_base->seq_cp(seq_id_src, seq_id_dst, p0, p1);
     kv_swa ->seq_cp(seq_id_src, seq_id_dst, p0, p1);
 }
 
-void llama_kv_cache_unified_iswa::seq_keep(llama_seq_id seq_id) {
+void llama_kv_cache_iswa::seq_keep(llama_seq_id seq_id) {
     kv_base->seq_keep(seq_id);
     kv_swa ->seq_keep(seq_id);
 }
 
-void llama_kv_cache_unified_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+void llama_kv_cache_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     kv_base->seq_add(seq_id, p0, p1, shift);
     kv_swa ->seq_add(seq_id, p0, p1, shift);
 }
 
-void llama_kv_cache_unified_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+void llama_kv_cache_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     kv_base->seq_div(seq_id, p0, p1, d);
     kv_swa ->seq_div(seq_id, p0, p1, d);
 }
 
-llama_pos llama_kv_cache_unified_iswa::seq_pos_min(llama_seq_id seq_id) const {
+llama_pos llama_kv_cache_iswa::seq_pos_min(llama_seq_id seq_id) const {
     // the base cache is a superset of the SWA cache, so we can just check the SWA cache
     return kv_swa->seq_pos_min(seq_id);
 }
 
-llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const {
+llama_pos llama_kv_cache_iswa::seq_pos_max(llama_seq_id seq_id) const {
     return kv_swa->seq_pos_max(seq_id);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+llama_memory_context_ptr llama_kv_cache_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
     GGML_UNUSED(embd_all);
 
     // first try simple split
@@ -136,7 +153,7 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
         assert(sinfos_base.size() == sinfos_swa.size());
 
-        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+        return std::make_unique<llama_kv_cache_iswa_context>(
                 this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
@@ -172,61 +189,67 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
         assert(sinfos_base.size() == sinfos_swa.size());
 
-        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+        return std::make_unique<llama_kv_cache_iswa_context>(
                 this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
     } while (false);
 
     // TODO: if we fail again, we should attempt different splitting strategies
     //       but to do that properly, we first have to refactor the batches to be more flexible
 
-    return std::make_unique<llama_kv_cache_unified_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+    return std::make_unique<llama_kv_cache_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified_iswa::init_full() {
-    return std::make_unique<llama_kv_cache_unified_iswa_context>(this);
+llama_memory_context_ptr llama_kv_cache_iswa::init_full() {
+    return std::make_unique<llama_kv_cache_iswa_context>(this);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) {
-    return std::make_unique<llama_kv_cache_unified_iswa_context>(this, lctx, optimize);
+llama_memory_context_ptr llama_kv_cache_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_iswa_context>(this, lctx, optimize);
 }
 
-bool llama_kv_cache_unified_iswa::get_can_shift() const {
+bool llama_kv_cache_iswa::get_can_shift() const {
     return kv_base->get_size() == kv_swa->get_size();
 }
 
-void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
-    kv_base->state_write(io, seq_id);
-    kv_swa ->state_write(io, seq_id);
+void llama_kv_cache_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
+        kv_base->state_write(io, seq_id, flags);
+    }
+
+    kv_swa->state_write(io, seq_id, flags);
 }
 
-void llama_kv_cache_unified_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
-    kv_base->state_read(io, seq_id);
-    kv_swa ->state_read(io, seq_id);
+void llama_kv_cache_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
+        kv_base->state_read(io, seq_id, flags);
+    }
+
+    kv_swa->state_read(io, seq_id, flags);
 }
 
-llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_base() const {
+llama_kv_cache * llama_kv_cache_iswa::get_base() const {
     return kv_base.get();
 }
 
-llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_swa() const {
+llama_kv_cache * llama_kv_cache_iswa::get_swa() const {
     return kv_swa.get();
 }
 
 //
-// llama_kv_cache_unified_iswa_context
+// llama_kv_cache_iswa_context
 //
 
-llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(llama_memory_status status) : status(status) {}
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(llama_memory_status status) : status(status) {}
 
-llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
-        llama_kv_cache_unified_iswa * kv) :
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv) :
     ctx_base(kv->get_base()->init_full()),
     ctx_swa (kv->get_swa ()->init_full()),
     status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
-        llama_kv_cache_unified_iswa * kv,
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv,
         llama_context * lctx,
         bool optimize) :
     ctx_base(kv->get_base()->init_update(lctx, optimize)),
@@ -234,21 +257,21 @@ llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
     status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
-        llama_kv_cache_unified_iswa * kv,
+llama_kv_cache_iswa_context::llama_kv_cache_iswa_context(
+        llama_kv_cache_iswa * kv,
         slot_info_vec_t sinfos_base,
         slot_info_vec_t sinfos_swa,
         std::vector<llama_ubatch> ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
-    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
+    ctx_base(new llama_kv_cache_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
+    ctx_swa (new llama_kv_cache_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
     status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_context:: ~llama_kv_cache_unified_iswa_context() = default;
+llama_kv_cache_iswa_context:: ~llama_kv_cache_iswa_context() = default;
 
-bool llama_kv_cache_unified_iswa_context::next() {
+bool llama_kv_cache_iswa_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     ctx_base->next();
@@ -261,7 +284,7 @@ bool llama_kv_cache_unified_iswa_context::next() {
     return true;
 }
 
-bool llama_kv_cache_unified_iswa_context::apply() {
+bool llama_kv_cache_iswa_context::apply() {
     assert(!llama_memory_status_is_fail(status));
 
     bool res = true;
@@ -272,24 +295,24 @@ bool llama_kv_cache_unified_iswa_context::apply() {
     return res;
 }
 
-llama_memory_status llama_kv_cache_unified_iswa_context::get_status() const {
+llama_memory_status llama_kv_cache_iswa_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_kv_cache_unified_iswa_context::get_ubatch() const {
+const llama_ubatch & llama_kv_cache_iswa_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     return ubatches[i_next];
 }
 
-const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_base() const {
+const llama_kv_cache_context * llama_kv_cache_iswa_context::get_base() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return static_cast<const llama_kv_cache_unified_context *>(ctx_base.get());
+    return static_cast<const llama_kv_cache_context *>(ctx_base.get());
 }
 
-const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_swa()  const {
+const llama_kv_cache_context * llama_kv_cache_iswa_context::get_swa()  const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return static_cast<const llama_kv_cache_unified_context *>(ctx_swa.get());
+    return static_cast<const llama_kv_cache_context *>(ctx_swa.get());
 }
diff --git a/src/llama-kv-cache-unified-iswa.h b/src/llama-kv-cache-iswa.h
similarity index 67%
rename from src/llama-kv-cache-unified-iswa.h
rename to src/llama-kv-cache-iswa.h
index d2650dadd3595..5ed134b795800 100644
--- a/src/llama-kv-cache-unified-iswa.h
+++ b/src/llama-kv-cache-iswa.h
@@ -1,19 +1,19 @@
 #pragma once
 
-#include "llama-kv-cache-unified.h"
+#include "llama-kv-cache.h"
 
 #include <vector>
 
 //
-// llama_kv_cache_unified_iswa
+// llama_kv_cache_iswa
 //
 
-// utilizes two instances of llama_kv_cache_unified
+// utilizes two instances of llama_kv_cache
 //   the first instance is for the non-SWA layers of the model and the second instance is for the SWA layers
 
-class llama_kv_cache_unified_iswa : public llama_memory_i {
+class llama_kv_cache_iswa : public llama_memory_i {
 public:
-    llama_kv_cache_unified_iswa(
+    llama_kv_cache_iswa(
             const llama_model & model,
                     ggml_type   type_k,
                     ggml_type   type_v,
@@ -24,9 +24,11 @@ class llama_kv_cache_unified_iswa : public llama_memory_i {
                      uint32_t   kv_size,
                      uint32_t   n_seq_max,
                      uint32_t   n_ubatch,
-                     uint32_t   n_pad);
+                     uint32_t   n_pad,
+        const layer_filter_cb & filter,
+        const  layer_reuse_cb & reuse);
 
-    ~llama_kv_cache_unified_iswa() = default;
+    ~llama_kv_cache_iswa() = default;
 
     //
     // llama_memory_i
@@ -56,50 +58,50 @@ class llama_kv_cache_unified_iswa : public llama_memory_i {
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     //
-    // llama_kv_cache_unified_iswa specific API
+    // llama_kv_cache_iswa specific API
     //
 
-    llama_kv_cache_unified * get_base() const;
-    llama_kv_cache_unified * get_swa () const;
+    llama_kv_cache * get_base() const;
+    llama_kv_cache * get_swa () const;
 
 private:
     const llama_hparams & hparams;
 
     const bool unified;
 
-    std::unique_ptr<llama_kv_cache_unified> kv_base;
-    std::unique_ptr<llama_kv_cache_unified> kv_swa;
+    std::unique_ptr<llama_kv_cache> kv_base;
+    std::unique_ptr<llama_kv_cache> kv_swa;
 };
 
-class llama_kv_cache_unified_iswa_context : public llama_memory_context_i {
+class llama_kv_cache_iswa_context : public llama_memory_context_i {
 public:
-    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+    using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
 
     // used for errors
-    llama_kv_cache_unified_iswa_context(llama_memory_status status);
+    llama_kv_cache_iswa_context(llama_memory_status status);
 
     // used to create a full-cache context
-    llama_kv_cache_unified_iswa_context(
-            llama_kv_cache_unified_iswa * kv);
+    llama_kv_cache_iswa_context(
+            llama_kv_cache_iswa * kv);
 
     // used to create an update context
-    llama_kv_cache_unified_iswa_context(
-            llama_kv_cache_unified_iswa * kv,
+    llama_kv_cache_iswa_context(
+            llama_kv_cache_iswa * kv,
             llama_context * lctx,
             bool optimize);
 
     // used to create a batch processing context from a batch
-    llama_kv_cache_unified_iswa_context(
-            llama_kv_cache_unified_iswa * kv,
+    llama_kv_cache_iswa_context(
+            llama_kv_cache_iswa * kv,
             slot_info_vec_t sinfos_base,
             slot_info_vec_t sinfos_swa,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_unified_iswa_context();
+    virtual ~llama_kv_cache_iswa_context();
 
     //
     // llama_memory_context_i
@@ -112,14 +114,14 @@ class llama_kv_cache_unified_iswa_context : public llama_memory_context_i {
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_unified_iswa_context specific API
+    // llama_kv_cache_iswa_context specific API
     //
 
-    const llama_kv_cache_unified_context * get_base() const;
-    const llama_kv_cache_unified_context * get_swa()  const;
+    const llama_kv_cache_context * get_base() const;
+    const llama_kv_cache_context * get_swa()  const;
 
 private:
-    //llama_kv_cache_unified_iswa * kv;
+    //llama_kv_cache_iswa * kv;
 
     // the index of the next ubatch to process
     size_t i_next = 0;
diff --git a/src/llama-kv-cache-unified.cpp b/src/llama-kv-cache.cpp
similarity index 72%
rename from src/llama-kv-cache-unified.cpp
rename to src/llama-kv-cache.cpp
index e539142e6b8cd..d7ab56ccd9aac 100644
--- a/src/llama-kv-cache-unified.cpp
+++ b/src/llama-kv-cache.cpp
@@ -1,4 +1,4 @@
-#include "llama-kv-cache-unified.h"
+#include "llama-kv-cache.h"
 
 #include "llama-impl.h"
 #include "llama-io.h"
@@ -13,36 +13,29 @@
 #include <stdexcept>
 
 //
-// llama_kv_cache_unified
+// llama_kv_cache
 //
 
-llama_kv_cache_unified::llama_kv_cache_unified(
-        const llama_model &  model,
-          layer_filter_cb && filter,
-                ggml_type    type_k,
-                ggml_type    type_v,
-                     bool    v_trans,
-                     bool    offload,
-                     bool    unified,
-                 uint32_t    kv_size,
-                 uint32_t    n_seq_max,
-                 uint32_t    n_pad,
-                 uint32_t    n_swa,
-           llama_swa_type    swa_type) :
+llama_kv_cache::llama_kv_cache(
+        const llama_model & model,
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   offload,
+                     bool   unified,
+                 uint32_t   kv_size,
+                 uint32_t   n_seq_max,
+                 uint32_t   n_pad,
+                 uint32_t   n_swa,
+           llama_swa_type   swa_type,
+    const layer_filter_cb & filter,
+    const  layer_reuse_cb & reuse) :
     model(model), hparams(model.hparams), v_trans(v_trans),
     n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa), swa_type(swa_type) {
 
     GGML_ASSERT(kv_size % n_pad == 0);
 
-    // TODO: this is temporary until we support passing reuse layer filters [KV_REUSE]
-    auto n_layer_cache = hparams.n_layer;
-    if (model.arch == LLM_ARCH_GEMMA3N) {
-        n_layer_cache = 20;
-    }
-    if (model.arch == LLM_ARCH_GLM4_MOE) {
-        // GLM-4.5: Only process up to last layer, skip final NextN layer
-        n_layer_cache = hparams.n_layer - hparams.nextn_predict_layers;
-    }
+    const uint32_t n_layer_kv = hparams.n_layer_kv();
 
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
@@ -50,7 +43,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
             ggml_init_params params = {
-                /*.mem_size   =*/ size_t(2u*(1 + n_stream)*n_layer_cache*ggml_tensor_overhead()),
+                /*.mem_size   =*/ size_t(2u*(1 + n_stream)*n_layer_kv*ggml_tensor_overhead()),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
             };
@@ -97,9 +90,14 @@ llama_kv_cache_unified::llama_kv_cache_unified(
                 __func__, hparams.n_embd_v_gqa_max());
     }
 
-    for (uint32_t il = 0; il < n_layer_cache; il++) {
+    for (uint32_t il = 0; il < hparams.n_layer; il++) {
+        if (!hparams.has_kv(il)) {
+            LLAMA_LOG_DEBUG("%s: layer %3d: does not have KV cache\n", __func__, il);
+            continue;
+        }
+
         if (filter && !filter(il)) {
-            LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, il);
+            LLAMA_LOG_DEBUG("%s: layer %3d: filtered\n", __func__, il);
             continue;
         }
 
@@ -147,23 +145,27 @@ llama_kv_cache_unified::llama_kv_cache_unified(
         layers.push_back({ il, k, v, k_stream, v_stream, });
     }
 
-    // TODO: this is temporary until we support passing reuse layer filters [KV_REUSE]
-    if (model.arch == LLM_ARCH_GEMMA3N) {
-        LLAMA_LOG_DEBUG("%s: GEMMA3N: reuse layers [%d, %d]\n", __func__, n_layer_cache, hparams.n_layer - 1);
+    if (reuse) {
+        LLAMA_LOG_DEBUG("%s: reusing layers:\n", __func__);
 
-        for (uint32_t il = n_layer_cache; il < hparams.n_layer; il++) {
-            if (filter && !filter(il)) {
-                LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, il);
+        for (uint32_t il = 0; il < hparams.n_layer; il++) {
+            const int32_t il_reuse = reuse(il);
+
+            if (il_reuse < 0) {
+                LLAMA_LOG_DEBUG("%s: - layer %3d: no reuse\n", __func__, il);
                 continue;
             }
 
-            const bool     is_swa   = hparams.is_swa(il);
-            const uint32_t il_reuse = n_layer_cache - (is_swa ? 2 : 1);
+            if (filter && !filter(il)) {
+                LLAMA_LOG_DEBUG("%s: - layer %3d: filtered\n", __func__, il);
+                continue;
+            }
 
             GGML_ASSERT(map_layer_ids.find(il_reuse) != map_layer_ids.end());
+
             map_layer_ids[il] = map_layer_ids[il_reuse];
 
-            LLAMA_LOG_DEBUG("%s: layer %3d: reuse layer %d, isw = %d\n", __func__, il, il_reuse, is_swa);
+            LLAMA_LOG_DEBUG("%s: - layer %3d: reuse layer %d, is_swa = %d\n", __func__, il, il_reuse, hparams.is_swa(il));
         }
     }
 
@@ -209,7 +211,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     }
 }
 
-void llama_kv_cache_unified::clear(bool data) {
+void llama_kv_cache::clear(bool data) {
     for (uint32_t s = 0; s < n_stream; ++s) {
         v_cells[s].reset();
         v_heads[s] = 0;
@@ -222,13 +224,8 @@ void llama_kv_cache_unified::clear(bool data) {
     }
 }
 
-bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
-    GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
-
-    auto & cells = v_cells[seq_to_stream[seq_id]];
-    auto & head  = v_heads[seq_to_stream[seq_id]];
-
-    uint32_t new_head = cells.size();
+bool llama_kv_cache::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     if (p0 < 0) {
         p0 = 0;
@@ -239,6 +236,11 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
     }
 
     if (seq_id >= 0) {
+        auto & cells = v_cells[seq_to_stream[seq_id]];
+        auto & head  = v_heads[seq_to_stream[seq_id]];
+
+        uint32_t new_head = cells.size();
+
         for (uint32_t i = 0; i < cells.size(); ++i) {
             if (!cells.pos_in(i, p0, p1)) {
                 continue;
@@ -250,30 +252,42 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
                 }
             }
         }
+
+        // If we freed up a slot, set head to it so searching can start there.
+        if (new_head != cells.size() && new_head < head) {
+            head = new_head;
+        }
     } else {
         // match any sequence
-        for (uint32_t i = 0; i < cells.size(); ++i) {
-            if (!cells.pos_in(i, p0, p1)) {
-                continue;
-            }
+        for (uint32_t s = 0; s < n_stream; ++s) {
+            auto & cells = v_cells[s];
+            auto & head  = v_heads[s];
 
-            cells.rm(i);
+            uint32_t new_head = cells.size();
 
-            if (new_head == cells.size()) {
-                new_head = i;
+            for (uint32_t i = 0; i < cells.size(); ++i) {
+                if (!cells.pos_in(i, p0, p1)) {
+                    continue;
+                }
+
+                cells.rm(i);
+
+                if (new_head == cells.size()) {
+                    new_head = i;
+                }
             }
-        }
-    }
 
-    // If we freed up a slot, set head to it so searching can start there.
-    if (new_head != cells.size() && new_head < head) {
-        head = new_head;
+            // If we freed up a slot, set head to it so searching can start there.
+            if (new_head != cells.size() && new_head < head) {
+                head = new_head;
+            }
+        }
     }
 
     return true;
 }
 
-void llama_kv_cache_unified::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     GGML_ASSERT(seq_id_src >= 0 && (size_t) seq_id_src < seq_to_stream.size());
     GGML_ASSERT(seq_id_dst >= 0 && (size_t) seq_id_dst < seq_to_stream.size());
 
@@ -356,7 +370,7 @@ void llama_kv_cache_unified::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id
     //}
 }
 
-void llama_kv_cache_unified::seq_keep(llama_seq_id seq_id) {
+void llama_kv_cache::seq_keep(llama_seq_id seq_id) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
@@ -378,7 +392,7 @@ void llama_kv_cache_unified::seq_keep(llama_seq_id seq_id) {
     }
 }
 
-void llama_kv_cache_unified::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
@@ -422,7 +436,7 @@ void llama_kv_cache_unified::seq_add(llama_seq_id seq_id, llama_pos p0, llama_po
     head = new_head != cells.size() ? new_head : 0;
 }
 
-void llama_kv_cache_unified::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+void llama_kv_cache::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
 
     auto & cells = v_cells[seq_to_stream[seq_id]];
@@ -455,7 +469,7 @@ void llama_kv_cache_unified::seq_div(llama_seq_id seq_id, llama_pos p0, llama_po
     }
 }
 
-llama_pos llama_kv_cache_unified::seq_pos_min(llama_seq_id seq_id) const {
+llama_pos llama_kv_cache::seq_pos_min(llama_seq_id seq_id) const {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
 
     const auto & cells = v_cells[seq_to_stream[seq_id]];
@@ -463,7 +477,7 @@ llama_pos llama_kv_cache_unified::seq_pos_min(llama_seq_id seq_id) const {
     return cells.seq_pos_min(seq_id);
 }
 
-llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const {
+llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
     GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
 
     const auto & cells = v_cells[seq_to_stream[seq_id]];
@@ -471,7 +485,7 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const {
     return cells.seq_pos_max(seq_id);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified::init_batch(
+llama_memory_context_ptr llama_kv_cache::init_batch(
             llama_batch_allocr & balloc,
             uint32_t n_ubatch,
             bool embd_all) {
@@ -501,62 +515,34 @@ llama_memory_context_ptr llama_kv_cache_unified::init_batch(
             break;
         }
 
-        return std::make_unique<llama_kv_cache_unified_context>(
+        return std::make_unique<llama_kv_cache_context>(
                 this, std::move(sinfos), std::move(ubatches));
     } while (false);
 
-    return std::make_unique<llama_kv_cache_unified_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+    return std::make_unique<llama_kv_cache_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified::init_full() {
-    return std::make_unique<llama_kv_cache_unified_context>(this);
+llama_memory_context_ptr llama_kv_cache::init_full() {
+    return std::make_unique<llama_kv_cache_context>(this);
 }
 
-llama_memory_context_ptr llama_kv_cache_unified::init_update(llama_context * lctx, bool optimize) {
-    bool do_shift = get_has_shift();
-
-    defrag_info dinfo;
-
-    // see if we need to defrag
-    if (n_stream == 1) {
-        // note : for now do not consider defrag for n_stream > 1
-        const auto & cells = v_cells[seq_to_stream[0]];
-
-        bool do_defrag = optimize;
-
-        const auto thold = lctx->get_cparams().defrag_thold;
+llama_memory_context_ptr llama_kv_cache::init_update(llama_context * lctx, bool optimize) {
+    GGML_UNUSED(optimize);
 
-        if (!do_defrag && thold > 0.0f) {
-            const auto n_kv = cells.used_max_p1();
-
-            // - do not defrag small contexts (i.e. < 2048 tokens)
-            // - count the padding towards the number of used tokens
-            const float fragmentation = n_kv >= 2048 ? std::max(0.0f, 1.0f - (float(cells.get_used() + n_pad)/n_kv)) : 0.0f;
-
-            if (fragmentation > thold) {
-                LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
-
-                do_defrag = true;
-            }
-        }
-
-        if (do_defrag) {
-            dinfo = defrag_prepare(lctx->graph_max_nodes());
-        }
-    }
+    bool do_shift = get_has_shift();
 
-    return std::make_unique<llama_kv_cache_unified_context>(this, lctx, do_shift, std::move(dinfo), std::move(sc_info));
+    return std::make_unique<llama_kv_cache_context>(this, lctx, do_shift, std::move(sc_info));
 }
 
-llama_kv_cache_unified::slot_info_vec_t llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
-    llama_kv_cache_unified::slot_info_vec_t res;
+llama_kv_cache::slot_info_vec_t llama_kv_cache::prepare(const std::vector<llama_ubatch> & ubatches) {
+    llama_kv_cache::slot_info_vec_t res;
 
     struct state_t {
         slot_info sinfo; // slot info for the ubatch
 
         std::vector<uint32_t> v_heads_old; // old positions of the heads, before placing the ubatch
 
-        std::vector<llama_kv_cells_unified> v_cells; // copy of the old cells, before placing the ubatch
+        std::vector<llama_kv_cells> v_cells; // copy of the old cells, before placing the ubatch
     };
 
     // remember the old state of the cells so we can restore it in the end
@@ -617,7 +603,7 @@ llama_kv_cache_unified::slot_info_vec_t llama_kv_cache_unified::prepare(const st
     return res;
 }
 
-bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const defrag_info & dinfo, const stream_copy_info & sc_info) {
+bool llama_kv_cache::update(llama_context * lctx, bool do_shift, const stream_copy_info & sc_info) {
     bool updated = false;
 
     auto * sched = lctx->get_sched();
@@ -687,117 +673,74 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
         }
     }
 
-    if (!dinfo.empty()) {
-        LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__);
-
-        // note: for now do not consider defrag for n_stream > 1
-        auto & cells = v_cells[seq_to_stream[0]];
-        auto & head  = v_heads[seq_to_stream[0]];
-
-        // apply moves:
-        {
-            const auto n_kv = dinfo.ids.size();
-
-            for (uint32_t i = 0; i < n_kv; ++i) {
-                assert(dinfo.ids[i] <= n_kv);
-
-                if (dinfo.ids[i] == n_kv || dinfo.ids[i] == i) {
-                    continue;
-                }
-
-                cells.mv(i, dinfo.ids[i]);
-            }
-
-            // reset the head so we can find the first free slot during the next ubatch
-            head = 0;
-        }
-
-        ggml_backend_sched_reset(sched);
-
-        auto * res = lctx->get_gf_res_reserve();
-
-        res->reset();
-
-        auto * gf = build_graph_defrag(res, lctx, dinfo);
-        if (!ggml_backend_sched_alloc_graph(sched, gf)) {
-            LLAMA_LOG_ERROR("%s: failed to allocate compute graph for defrag\n", __func__);
-            return updated;
-        }
-
-        res->set_inputs(nullptr);
-
-        if (lctx->graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
-            LLAMA_LOG_ERROR("%s: failed to compute defrag\n", __func__);
-            return updated;
-        }
-
-        updated = true;
-    }
-
     return updated;
 }
 
-llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch, bool cont) const {
-    if (debug > 0) {
-        const auto & cells = v_cells[seq_to_stream[1]];
-
-        const uint32_t head_cur = v_heads[1];
+llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch, bool cont) const {
 
-        LLAMA_LOG_DEBUG("%s: n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n",
-                __func__, cells.used_max_p1(), cells.get_used(), head_cur, get_size(), n_swa);
+    if (debug > 0) {
+        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+            const auto seq_id = ubatch.seq_id_unq[s];
+            const auto stream_id = seq_to_stream[seq_id];
+            const auto & cells = v_cells[stream_id];
+            const uint32_t head_cur = v_heads[stream_id];
+
+            LLAMA_LOG_DEBUG("%s: stream[%d], n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n",
+                    __func__, stream_id, cells.used_max_p1(), cells.get_used(), head_cur, get_size(), n_swa);
+
+            if ((debug == 2 && n_swa > 0) || debug > 2) {
+                std::string ss;
+                for (uint32_t i = 0; i < cells.size(); ++i) {
+                    if (cells.is_empty(i)) {
+                        ss += '.';
+                    } else {
+                        assert(cells.seq_count(i) >= 1);
 
-        if ((debug == 2 && n_swa > 0) || debug > 2) {
-            std::string ss;
-            for (uint32_t i = 0; i < cells.size(); ++i) {
-                if (cells.is_empty(i)) {
-                    ss += '.';
-                } else {
-                    assert(cells.seq_count(i) >= 1);
+                        if (cells.seq_count(i) == 1) {
+                            ss += std::to_string(cells.seq_get(i));
+                        } else {
+                            ss += 'M';
+                        }
+                    }
+                    if (i%256 == 255) {
+                        ss += " *";
+                        ss += '\n';
+                    }
+                }
+                LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
+            }
 
-                    if (cells.seq_count(i) == 1) {
-                        ss += std::to_string(cells.seq_get(i));
+            if ((debug == 2 && n_swa > 0) || debug > 2) {
+                std::string ss;
+                for (uint32_t i = 0; i < cells.size(); ++i) {
+                    std::string cur;
+                    if (cells.is_empty(i)) {
+                        cur = '.';
                     } else {
-                        ss += 'M';
+                        cur = std::to_string(cells.pos_get(i));
+                    }
+                    const int n = cur.size();
+                    for (int j = 0; j < 5 - n; ++j) {
+                        cur += ' ';
+                    }
+                    ss += cur;
+                    if (i%256 == 255) {
+                        ss += " *";
+                    }
+                    if (i%64 == 63) {
+                        ss += '\n';
                     }
                 }
-                if (i%256 == 255) {
-                    ss += " *";
-                    ss += '\n';
-                }
+                LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
             }
-            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
-        }
 
-        if ((debug == 2 && n_swa > 0) || debug > 2) {
-            std::string ss;
-            for (uint32_t i = 0; i < cells.size(); ++i) {
-                std::string cur;
-                if (cells.is_empty(i)) {
-                    cur = '.';
-                } else {
-                    cur = std::to_string(cells.pos_get(i));
-                }
-                const int n = cur.size();
-                for (int j = 0; j < 5 - n; ++j) {
-                    cur += ' ';
-                }
-                ss += cur;
-                if (i%256 == 255) {
-                    ss += " *";
-                }
-                if (i%64 == 63) {
-                    ss += '\n';
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+                if (cells.seq_pos_min(s) < 0) {
+                    continue;
                 }
-            }
-            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
-        }
 
-        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
-            if (cells.seq_pos_min(s) < 0) {
-                continue;
+                LLAMA_LOG_DEBUG("%s: stream[%d] min[%d] = %5d, max[%d] = %5d\n", __func__, stream_id, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
             }
-
-            LLAMA_LOG_DEBUG("%s: min[%d] = %5d, max[%d] = %5d\n", __func__, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
         }
     }
 
@@ -932,7 +875,7 @@ llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_
     return res;
 }
 
-void llama_kv_cache_unified::apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch) {
+void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch) {
     // keep track of the max sequence position that we would overwrite with this ubatch
     // for non-SWA cache, this would be always empty
     llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ];
@@ -997,21 +940,21 @@ void llama_kv_cache_unified::apply_ubatch(const slot_info & sinfo, const llama_u
     }
 }
 
-bool llama_kv_cache_unified::get_can_shift() const {
+bool llama_kv_cache::get_can_shift() const {
     return true;
 }
 
-uint32_t llama_kv_cache_unified::get_size() const {
+uint32_t llama_kv_cache::get_size() const {
     const auto & cells = v_cells[seq_to_stream[0]];
 
     return cells.size();
 }
 
-uint32_t llama_kv_cache_unified::get_n_stream() const {
+uint32_t llama_kv_cache::get_n_stream() const {
     return n_stream;
 }
 
-bool llama_kv_cache_unified::get_has_shift() const {
+bool llama_kv_cache::get_has_shift() const {
     bool result = false;
 
     for (uint32_t s = 0; s < n_stream; ++s) {
@@ -1021,7 +964,7 @@ bool llama_kv_cache_unified::get_has_shift() const {
     return result;
 }
 
-uint32_t llama_kv_cache_unified::get_n_kv() const {
+uint32_t llama_kv_cache::get_n_kv() const {
     uint32_t result = 0;
 
     for (uint32_t s = 0; s < n_stream; ++s) {
@@ -1033,11 +976,11 @@ uint32_t llama_kv_cache_unified::get_n_kv() const {
     return result;
 }
 
-bool llama_kv_cache_unified::get_supports_set_rows() const {
+bool llama_kv_cache::get_supports_set_rows() const {
     return supports_set_rows;
 }
 
-ggml_tensor * llama_kv_cache_unified::get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const {
+ggml_tensor * llama_kv_cache::get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * k = layers[ikv].k;
@@ -1057,7 +1000,7 @@ ggml_tensor * llama_kv_cache_unified::get_k(ggml_context * ctx, int32_t il, uint
             ggml_row_size(k->type, n_embd_k_gqa*kv_size)*sinfo.s0);
 }
 
-ggml_tensor * llama_kv_cache_unified::get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const {
+ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * v = layers[ikv].v;
@@ -1089,7 +1032,7 @@ ggml_tensor * llama_kv_cache_unified::get_v(ggml_context * ctx, int32_t il, uint
             ggml_row_size(v->type, kv_size*n_embd_v_gqa)*sinfo.s0);
 }
 
-ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const {
+ggml_tensor * llama_kv_cache::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * k = layers[ikv].k;
@@ -1119,7 +1062,7 @@ ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_
     return ggml_cpy(ctx, k_cur, k_view);
 }
 
-ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const {
+ggml_tensor * llama_kv_cache::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * v = layers[ikv].v;
@@ -1173,7 +1116,7 @@ ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_
     return ggml_cpy(ctx, v_cur, v_view);
 }
 
-ggml_tensor * llama_kv_cache_unified::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+ggml_tensor * llama_kv_cache::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
     const uint32_t n_tokens = ubatch.n_tokens;
 
     ggml_tensor * k_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
@@ -1183,7 +1126,7 @@ ggml_tensor * llama_kv_cache_unified::build_input_k_idxs(ggml_context * ctx, con
     return k_idxs;
 }
 
-ggml_tensor * llama_kv_cache_unified::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+ggml_tensor * llama_kv_cache::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
     const uint32_t n_tokens = ubatch.n_tokens;
 
     ggml_tensor * v_idxs;
@@ -1199,7 +1142,7 @@ ggml_tensor * llama_kv_cache_unified::build_input_v_idxs(ggml_context * ctx, con
     return v_idxs;
 }
 
-void llama_kv_cache_unified::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
+void llama_kv_cache::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
     if (!supports_set_rows) {
         return;
     }
@@ -1219,7 +1162,7 @@ void llama_kv_cache_unified::set_input_k_idxs(ggml_tensor * dst, const llama_uba
     }
 }
 
-void llama_kv_cache_unified::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
+void llama_kv_cache::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
     if (!supports_set_rows) {
         return;
     }
@@ -1256,7 +1199,7 @@ void llama_kv_cache_unified::set_input_v_idxs(ggml_tensor * dst, const llama_uba
     }
 }
 
-void llama_kv_cache_unified::set_input_k_shift(ggml_tensor * dst) const {
+void llama_kv_cache::set_input_k_shift(ggml_tensor * dst) const {
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
 
     int32_t * data = (int32_t *) dst->data;
@@ -1270,7 +1213,7 @@ void llama_kv_cache_unified::set_input_k_shift(ggml_tensor * dst) const {
     }
 }
 
-void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
+void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
     const uint32_t n_tokens = ubatch->n_tokens;
 
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
@@ -1342,7 +1285,7 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
     }
 }
 
-void llama_kv_cache_unified::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     const int64_t n_tokens = ubatch->n_tokens;
 
     GGML_ASSERT(n_stream == 1 && "TODO: support multiple streams");
@@ -1367,7 +1310,7 @@ void llama_kv_cache_unified::set_input_pos_bucket(ggml_tensor * dst, const llama
     }
 }
 
-size_t llama_kv_cache_unified::total_size() const {
+size_t llama_kv_cache::total_size() const {
     size_t size = 0;
 
     for (const auto & buf : bufs) {
@@ -1377,7 +1320,7 @@ size_t llama_kv_cache_unified::total_size() const {
     return size;
 }
 
-size_t llama_kv_cache_unified::size_k_bytes() const {
+size_t llama_kv_cache::size_k_bytes() const {
     size_t size_k_bytes = 0;
 
     for (const auto & layer : layers) {
@@ -1387,7 +1330,7 @@ size_t llama_kv_cache_unified::size_k_bytes() const {
     return size_k_bytes;
 }
 
-size_t llama_kv_cache_unified::size_v_bytes() const {
+size_t llama_kv_cache::size_v_bytes() const {
     size_t size_v_bytes = 0;
 
     for (const auto & layer : layers) {
@@ -1397,7 +1340,7 @@ size_t llama_kv_cache_unified::size_v_bytes() const {
     return size_v_bytes;
 }
 
-ggml_tensor * llama_kv_cache_unified::build_rope_shift(
+ggml_tensor * llama_kv_cache::build_rope_shift(
         const llama_cparams & cparams,
                ggml_context * ctx,
                 ggml_tensor * cur,
@@ -1449,14 +1392,14 @@ ggml_tensor * llama_kv_cache_unified::build_rope_shift(
 
 class llm_graph_input_k_shift : public llm_graph_input_i {
 public:
-    llm_graph_input_k_shift(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    llm_graph_input_k_shift(const llama_kv_cache * kv_self) : kv_self(kv_self) {}
     virtual ~llm_graph_input_k_shift() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * k_shift; // I32 [kv_size*n_stream]
 
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache * kv_self;
 };
 
 void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) {
@@ -1467,7 +1410,7 @@ void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-ggml_cgraph * llama_kv_cache_unified::build_graph_shift(llm_graph_result * res, llama_context * lctx) const {
+ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_context * lctx) const {
     auto * ctx = res->get_ctx();
     auto * gf  = res->get_gf();
 
@@ -1509,284 +1452,7 @@ ggml_cgraph * llama_kv_cache_unified::build_graph_shift(llm_graph_result * res,
     return gf;
 }
 
-ggml_cgraph * llama_kv_cache_unified::build_graph_defrag(
-         llm_graph_result * res,
-            llama_context * lctx,
-        const defrag_info & dinfo) const {
-    auto * ctx = res->get_ctx();
-    auto * gf  = res->get_gf();
-
-    GGML_ASSERT(n_stream == 1 && "n_stream > 1 does not support defrag");
-
-    const auto & cells = v_cells[0];
-
-    const auto & ids = dinfo.ids;
-
-    const auto & cparams = lctx->get_cparams();
-
-#if 0
-    // CPU defrag
-    //
-    // TODO: optimizations are possible:
-    //       - multiple threads
-    //       - avoid copying to the host memory when already there
-    //
-    // likely not worth the effort, as we have ggml_graph based defrag
-    //
-
-    const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa();
-    const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa();
-
-    const uint32_t kv_size = size;
-
-    std::vector<uint8_t> buf_k;
-    std::vector<uint8_t> buf_v;
-
-    for (uint32_t il = 0; il < n_layer; ++il) {
-        const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa);
-        const size_t k_size     = ggml_row_size(k_l[il]->type, n_embd_k_gqa*kv_size);
-
-        const size_t v_size_el = ggml_type_size(v_l[il]->type);
-        const size_t v_size    = ggml_row_size (v_l[il]->type, n_embd_v_gqa*kv_size);
-
-        buf_k.resize(k_size);
-        buf_v.resize(v_size);
-
-        ggml_backend_tensor_get(k_l[il], buf_k.data(), 0, buf_k.size());
-        ggml_backend_tensor_get(v_l[il], buf_v.data(), 0, buf_v.size());
-
-        // batch move [i, i+nm) to [id, id+nm)
-        // note: cells can move only to a lower index
-        for (uint32_t i = 0; i < n_kv; ++i) {
-            const uint32_t id = ids[i];
-
-            if (i == id || id == n_kv) {
-                continue;
-            }
-
-            uint32_t nm = 1;
-
-            while (i + nm < n_kv && ids[i + nm] == id + nm) {
-                nm++;
-            }
-
-            // move keys
-            {
-                const int64_t os =  i*k_size_row;
-                const int64_t od = id*k_size_row;
-
-                memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row);
-            }
-
-            // move values (note: they are transposed)
-            {
-                const int64_t os =  i;
-                const int64_t od = id;
-
-                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                    memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el);
-                }
-            }
-
-            i += nm - 1;
-        }
-
-        ggml_backend_tensor_set(k_l[il], buf_k.data(), 0, buf_k.size());
-        ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size());
-    }
-#else
-    for (uint32_t i = 0; i < ids.size(); ++i) {
-        const uint32_t id = ids[i];
-
-        if (i == id || id == ids.size()) {
-            continue;
-        }
-
-        uint32_t nm = 1;
-
-        while (i + nm < ids.size() && ids[i + nm] == id + nm) {
-            nm++;
-        }
-
-        for (const auto & layer : layers) {
-            const uint32_t il = layer.il;
-
-            const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-            const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
-
-            ggml_tensor * view_k_src = ggml_view_2d(ctx, layer.k,
-                    n_embd_k_gqa, nm,
-                    ggml_row_size(layer.k->type, n_embd_k_gqa),
-                    ggml_row_size(layer.k->type, n_embd_k_gqa*i));
-
-            ggml_tensor * view_k_dst = ggml_view_2d(ctx, layer.k,
-                    n_embd_k_gqa, nm,
-                    ggml_row_size(layer.k->type, n_embd_k_gqa),
-                    ggml_row_size(layer.k->type, n_embd_k_gqa*id));
-
-            ggml_tensor * view_v_src;
-            ggml_tensor * view_v_dst;
-
-            if (cparams.flash_attn) {
-                // NOTE: the V cache is not transposed when using flash attention
-                view_v_src = ggml_view_2d(ctx, layer.v,
-                        n_embd_v_gqa, nm,
-                        ggml_row_size(layer.v->type, n_embd_v_gqa),
-                        ggml_row_size(layer.v->type, n_embd_v_gqa*i));
-
-                view_v_dst = ggml_view_2d(ctx, layer.v,
-                        n_embd_v_gqa, nm,
-                        ggml_row_size(layer.v->type, n_embd_v_gqa),
-                        ggml_row_size(layer.v->type, n_embd_v_gqa*id));
-            } else {
-                view_v_src = ggml_view_2d(ctx, layer.v,
-                        nm, n_embd_v_gqa,
-                        ggml_row_size(layer.v->type, cells.size()),
-                        ggml_row_size(layer.v->type, i));
-
-                view_v_dst = ggml_view_2d(ctx, layer.v,
-                        nm, n_embd_v_gqa,
-                        ggml_row_size(layer.v->type, cells.size()),
-                        ggml_row_size(layer.v->type, id));
-            }
-
-            ggml_build_forward_expand(gf, ggml_cpy(ctx, view_k_src, view_k_dst));
-            ggml_build_forward_expand(gf, ggml_cpy(ctx, view_v_src, view_v_dst));
-        }
-
-        i += nm - 1;
-    }
-
-    //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
-#endif
-
-    return gf;
-}
-
-llama_kv_cache_unified::defrag_info llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) const {
-    GGML_ASSERT(n_stream == 1 && "n_stream > 1 does not support defrag");
-
-    const auto & cells = v_cells[0];
-
-    const uint32_t n_layer = layers.size();
-
-    const uint32_t n_kv   = cells.used_max_p1();
-    const uint32_t n_used = cells.get_used();
-
-    assert(n_used <= n_kv);
-
-    //const int64_t t_start = ggml_time_us();
-
-    // number of cells moved
-    uint32_t n_moves = 0;
-
-    // each move requires 6*n_layer tensors (see graph_build_kv_self_defrag)
-    //   - source view, destination view, copy operation
-    //   - x2 for keys and values
-    //const uint32_t max_moves = max_nodes()/(6*n_layer);
-    // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516
-    const uint32_t max_moves = (n_max_nodes - 2*n_layer)/(6*n_layer);
-
-    // determine which KV cells to move where
-    defrag_info res;
-    auto & ids = res.ids;
-
-    ids.resize(n_kv, n_kv);
-
-    for (uint32_t i0 = 0; i0 < n_used; ++i0) {
-        if (!cells.is_empty(i0)) {
-            ids[i0] = i0;
-
-            continue;
-        }
-
-        // found a hole - fill it with data from the end of the cache
-
-        uint32_t nh = 1;
-
-        // determine the size of the hole
-        while (i0 + nh < n_used && cells.is_empty(i0 + nh)) {
-            nh++;
-        }
-
-        uint32_t nf = 0;
-        uint32_t is = n_kv - 1;
-
-        // starting from the end, find nh non-empty cells
-        for (; is > i0; --is) {
-            if (cells.is_empty(is) || ids[is] != n_kv) {
-                continue;
-            }
-
-            // non-empty cell which is not yet moved
-            nf++;
-
-            if (nf == nh) {
-                break;
-            }
-        }
-
-        // this can only happen if `n_used` is not accurate, which would be a bug
-        GGML_ASSERT(nf == nh && "KV defrag bug: nf != nh");
-
-        nf = 0;
-
-        uint32_t i1 = is;
-
-        // are we moving a continuous block of memory?
-        bool cont = false;
-
-        // should we stop searching for the next move?
-        bool stop = false;
-
-        // go back and move the nf cells to the hole
-        for (; i1 < n_kv; ++i1) {
-            if (cells.is_empty(i1) || ids[i1] != n_kv) {
-                if (n_moves == max_moves) {
-                    stop = true;
-                    break;
-                }
-
-                cont = false;
-                continue;
-            }
-
-            // this cell goes to (i0 + nf)
-            ids[i1] = i0 + nf;
-
-            if (!cont) {
-                n_moves++;
-                cont = true;
-            }
-
-            nf++;
-
-            if (nf == nh) {
-                break;
-            }
-        }
-
-        if (stop || n_moves == max_moves) {
-            break;
-        }
-
-        //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh);
-
-        i0 += nh - 1;
-    }
-
-    if (n_moves == 0) {
-        return {};
-    }
-
-    LLAMA_LOG_DEBUG("%s: (tmp log) KV defrag cell moves: %u\n", __func__, n_moves);
-
-    LLAMA_LOG_DEBUG("%s: expected gf nodes: %u\n", __func__, 6*n_moves*n_layer);
-
-    return res;
-}
-
-bool llama_kv_cache_unified::is_masked_swa(llama_pos p0, llama_pos p1) const {
+bool llama_kv_cache::is_masked_swa(llama_pos p0, llama_pos p1) const {
     assert(p0 >= 0 && p1 >= 0);
 
     switch (swa_type) {
@@ -1812,7 +1478,9 @@ bool llama_kv_cache_unified::is_masked_swa(llama_pos p0, llama_pos p1) const {
     return false;
 }
 
-void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_kv_cache::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     io.write(&n_stream, sizeof(n_stream));
 
     for (uint32_t s = 0; s < n_stream; ++s) {
@@ -1863,7 +1531,9 @@ void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq
     }
 }
 
-void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     uint32_t n_stream_cur;
@@ -1897,7 +1567,7 @@ void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_i
     }
 }
 
-void llama_kv_cache_unified::state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id) const {
+void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id) const {
     const auto & cells = v_cells[cr.strm];
 
     for (const auto & range : cr.data) {
@@ -1925,7 +1595,7 @@ void llama_kv_cache_unified::state_write_meta(llama_io_write_i & io, const cell_
     }
 }
 
-void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const {
+void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const {
     const auto & cells = v_cells[cr.strm];
 
     const uint32_t v_trans = this->v_trans ? 1 : 0;
@@ -2020,7 +1690,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const cell_
     }
 }
 
-bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id) {
+bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id) {
     auto & cells = v_cells[strm];
     auto & head  = v_heads[strm];
 
@@ -2117,7 +1787,7 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t strm
     return true;
 }
 
-bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count) {
+bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count) {
     auto & cells = v_cells[strm];
     auto & head  = v_heads[strm];
 
@@ -2254,13 +1924,13 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t strm
 }
 
 //
-// llama_kv_cache_unified_context
+// llama_kv_cache_context
 //
 
-llama_kv_cache_unified_context::llama_kv_cache_unified_context(llama_memory_status status) : status(status) {}
+llama_kv_cache_context::llama_kv_cache_context(llama_memory_status status) : status(status) {}
 
-llama_kv_cache_unified_context::llama_kv_cache_unified_context(
-        llama_kv_cache_unified * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv) {
+llama_kv_cache_context::llama_kv_cache_context(
+        llama_kv_cache * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv) {
     n_kv = kv->get_size();
 
     const uint32_t n_stream = kv->get_n_stream();
@@ -2276,26 +1946,25 @@ llama_kv_cache_unified_context::llama_kv_cache_unified_context(
     }
 }
 
-llama_kv_cache_unified_context::llama_kv_cache_unified_context(
-        llama_kv_cache_unified * kv,
+llama_kv_cache_context::llama_kv_cache_context(
+        llama_kv_cache * kv,
         llama_context * lctx,
         bool do_shift,
-        defrag_info dinfo,
-        stream_copy_info sc_info) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), lctx(lctx), do_shift(do_shift), dinfo(std::move(dinfo)), sc_info(std::move(sc_info)) {
-    if (!do_shift && this->dinfo.empty() && this->sc_info.empty()) {
+        stream_copy_info sc_info) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), lctx(lctx), do_shift(do_shift), sc_info(std::move(sc_info)) {
+    if (!do_shift && this->sc_info.empty()) {
         status = LLAMA_MEMORY_STATUS_NO_UPDATE;
     }
 }
 
-llama_kv_cache_unified_context::llama_kv_cache_unified_context(
-        llama_kv_cache_unified * kv,
-        llama_kv_cache_unified::slot_info_vec_t sinfos,
+llama_kv_cache_context::llama_kv_cache_context(
+        llama_kv_cache * kv,
+        llama_kv_cache::slot_info_vec_t sinfos,
         std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), sinfos(std::move(sinfos)), ubatches(std::move(ubatches)) {
 }
 
-llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default;
+llama_kv_cache_context::~llama_kv_cache_context() = default;
 
-bool llama_kv_cache_unified_context::next() {
+bool llama_kv_cache_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     if (++i_cur >= ubatches.size()) {
@@ -2305,12 +1974,12 @@ bool llama_kv_cache_unified_context::next() {
     return true;
 }
 
-bool llama_kv_cache_unified_context::apply() {
+bool llama_kv_cache_context::apply() {
     assert(!llama_memory_status_is_fail(status));
 
     // no ubatches -> this is a KV cache update
     if (ubatches.empty()) {
-        kv->update(lctx, do_shift, dinfo, sc_info);
+        kv->update(lctx, do_shift, sc_info);
 
         return true;
     }
@@ -2322,69 +1991,69 @@ bool llama_kv_cache_unified_context::apply() {
     return true;
 }
 
-llama_memory_status llama_kv_cache_unified_context::get_status() const {
+llama_memory_status llama_kv_cache_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_kv_cache_unified_context::get_ubatch() const {
+const llama_ubatch & llama_kv_cache_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     return ubatches[i_cur];
 }
 
-uint32_t llama_kv_cache_unified_context::get_n_kv() const {
+uint32_t llama_kv_cache_context::get_n_kv() const {
     return n_kv;
 }
 
-bool llama_kv_cache_unified_context::get_supports_set_rows() const {
+bool llama_kv_cache_context::get_supports_set_rows() const {
     return kv->get_supports_set_rows();
 }
 
-ggml_tensor * llama_kv_cache_unified_context::get_k(ggml_context * ctx, int32_t il) const {
+ggml_tensor * llama_kv_cache_context::get_k(ggml_context * ctx, int32_t il) const {
     return kv->get_k(ctx, il, n_kv, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::get_v(ggml_context * ctx, int32_t il) const {
+ggml_tensor * llama_kv_cache_context::get_v(ggml_context * ctx, int32_t il) const {
     return kv->get_v(ctx, il, n_kv, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const {
+ggml_tensor * llama_kv_cache_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const {
     return kv->cpy_k(ctx, k_cur, k_idxs, il, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const {
+ggml_tensor * llama_kv_cache_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const {
     return kv->cpy_v(ctx, v_cur, v_idxs, il, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+ggml_tensor * llama_kv_cache_context::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
     return kv->build_input_k_idxs(ctx, ubatch);
 }
 
-ggml_tensor * llama_kv_cache_unified_context::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+ggml_tensor * llama_kv_cache_context::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
     return kv->build_input_v_idxs(ctx, ubatch);
 }
 
-void llama_kv_cache_unified_context::set_input_k_shift(ggml_tensor * dst) const {
+void llama_kv_cache_context::set_input_k_shift(ggml_tensor * dst) const {
     kv->set_input_k_shift(dst);
 }
 
-void llama_kv_cache_unified_context::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+void llama_kv_cache_context::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_k_idxs(dst, ubatch, sinfos[i_cur]);
 }
 
-void llama_kv_cache_unified_context::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+void llama_kv_cache_context::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_v_idxs(dst, ubatch, sinfos[i_cur]);
 }
 
-void llama_kv_cache_unified_context::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
+void llama_kv_cache_context::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
     kv->set_input_kq_mask(dst, ubatch, causal_attn);
 }
 
-void llama_kv_cache_unified_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+void llama_kv_cache_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_pos_bucket(dst, ubatch);
 }
 
-uint32_t llama_kv_cache_unified::get_padding(const llama_cparams & cparams) {
+uint32_t llama_kv_cache::get_padding(const llama_cparams & cparams) {
     // the FA kernels require padding to avoid extra runtime boundary checks
     return cparams.flash_attn ? 256u : 32u;
 }
diff --git a/src/llama-kv-cache-unified.h b/src/llama-kv-cache.h
similarity index 80%
rename from src/llama-kv-cache-unified.h
rename to src/llama-kv-cache.h
index 342a675962e2a..76a5cb1e28e7e 100644
--- a/src/llama-kv-cache-unified.h
+++ b/src/llama-kv-cache.h
@@ -14,27 +14,13 @@ struct llama_model;
 struct llama_context;
 
 //
-// llama_kv_cache_unified
+// llama_kv_cache
 //
 
-class llama_kv_cache_unified : public llama_memory_i {
+class llama_kv_cache : public llama_memory_i {
 public:
     static uint32_t get_padding(const llama_cparams & cparams);
 
-    // this callback is used to filter out layers that should not be included in the cache
-    using layer_filter_cb = std::function<bool(int32_t il)>;
-
-    struct defrag_info {
-        bool empty() const {
-            return ids.empty();
-        }
-
-        // contains information about which cell moves where:
-        //  - cell i moves to ids[i]
-        //  - if ids[i] == i || ids[i] == ids.size(), then cell i is not moved
-        std::vector<uint32_t> ids;
-    };
-
     struct stream_copy_info {
         bool empty() const {
             assert(ssrc.size() == sdst.size());
@@ -92,21 +78,22 @@ class llama_kv_cache_unified : public llama_memory_i {
 
     using slot_info_vec_t = std::vector<slot_info>;
 
-    llama_kv_cache_unified(
-            const llama_model &  model,
-              layer_filter_cb && filter,
-                    ggml_type    type_k,
-                    ggml_type    type_v,
-                         bool    v_trans,
-                         bool    offload,
-                         bool    unified,
-                     uint32_t    kv_size,
-                     uint32_t    n_seq_max,
-                     uint32_t    n_pad,
-                     uint32_t    n_swa,
-               llama_swa_type    swa_type);
-
-    ~llama_kv_cache_unified() = default;
+    llama_kv_cache(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
+                         bool   unified,
+                     uint32_t   kv_size,
+                     uint32_t   n_seq_max,
+                     uint32_t   n_pad,
+                     uint32_t   n_swa,
+               llama_swa_type   swa_type,
+        const layer_filter_cb & filter,
+        const  layer_reuse_cb & reuse);
+
+    ~llama_kv_cache() = default;
 
     //
     // llama_memory_i
@@ -136,11 +123,11 @@ class llama_kv_cache_unified : public llama_memory_i {
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     //
-    // llama_kv_cache_unified specific API
+    // llama_kv_cache specific API
     //
 
     uint32_t get_size()     const;
@@ -173,7 +160,7 @@ class llama_kv_cache_unified : public llama_memory_i {
     // return empty vector on failure
     slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
 
-    bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo, const stream_copy_info & sc_info);
+    bool update(llama_context * lctx, bool do_shift, const stream_copy_info & sc_info);
 
     // find a slot of kv cells that can hold the ubatch
     // if cont == true, then the slot must be continuous
@@ -241,7 +228,7 @@ class llama_kv_cache_unified : public llama_memory_i {
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
     std::vector<uint32_t> v_heads;
 
-    std::vector<llama_kv_cells_unified> v_cells;
+    std::vector<llama_kv_cells> v_cells;
 
     // maps from a sequence id to a stream id
     std::vector<uint32_t> seq_to_stream;
@@ -254,9 +241,6 @@ class llama_kv_cache_unified : public llama_memory_i {
     // model layer id -> KV cache layer id
     std::unordered_map<int32_t, int32_t> map_layer_ids;
 
-    // return non-empty vector if cells have been moved
-    defrag_info defrag_prepare(int32_t n_max_nodes) const;
-
     size_t total_size() const;
 
     size_t size_k_bytes() const;
@@ -277,11 +261,6 @@ class llama_kv_cache_unified : public llama_memory_i {
                llm_graph_result * res,
                   llama_context * lctx) const;
 
-    ggml_cgraph * build_graph_defrag(
-               llm_graph_result * res,
-                  llama_context * lctx,
-              const defrag_info & dinfo) const;
-
     struct cell_ranges_t {
         uint32_t strm;
 
@@ -295,35 +274,33 @@ class llama_kv_cache_unified : public llama_memory_i {
     bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count);
 };
 
-class llama_kv_cache_unified_context : public llama_memory_context_i {
+class llama_kv_cache_context : public llama_memory_context_i {
 public:
     // some shorthands
-    using slot_info_vec_t  = llama_kv_cache_unified::slot_info_vec_t;
-    using defrag_info      = llama_kv_cache_unified::defrag_info;
-    using stream_copy_info = llama_kv_cache_unified::stream_copy_info;
+    using slot_info_vec_t  = llama_kv_cache::slot_info_vec_t;
+    using stream_copy_info = llama_kv_cache::stream_copy_info;
 
     // used for errors
-    llama_kv_cache_unified_context(llama_memory_status status);
+    llama_kv_cache_context(llama_memory_status status);
 
     // used to create a full-cache context
-    llama_kv_cache_unified_context(
-            llama_kv_cache_unified * kv);
+    llama_kv_cache_context(
+            llama_kv_cache * kv);
 
     // used to create an update context
-    llama_kv_cache_unified_context(
-            llama_kv_cache_unified * kv,
+    llama_kv_cache_context(
+            llama_kv_cache * kv,
             llama_context * lctx,
             bool do_shift,
-            defrag_info dinfo,
             stream_copy_info sc_info);
 
     // used to create a batch procesing context from a batch
-    llama_kv_cache_unified_context(
-            llama_kv_cache_unified * kv,
+    llama_kv_cache_context(
+            llama_kv_cache * kv,
             slot_info_vec_t sinfos,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_unified_context();
+    virtual ~llama_kv_cache_context();
 
     //
     // llama_memory_context_i
@@ -336,7 +313,7 @@ class llama_kv_cache_unified_context : public llama_memory_context_i {
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_unified_context specific API
+    // llama_kv_cache_context specific API
     //
 
     uint32_t get_n_kv() const;
@@ -365,7 +342,7 @@ class llama_kv_cache_unified_context : public llama_memory_context_i {
 private:
     llama_memory_status status;
 
-    llama_kv_cache_unified * kv;
+    llama_kv_cache * kv;
     llama_context * lctx;
 
     //
@@ -374,8 +351,6 @@ class llama_kv_cache_unified_context : public llama_memory_context_i {
 
     bool do_shift = false;
 
-    defrag_info dinfo;
-
     stream_copy_info sc_info;
 
     //
diff --git a/src/llama-kv-cells.h b/src/llama-kv-cells.h
index 0d0dd316fd041..8f6bf01456c8f 100644
--- a/src/llama-kv-cells.h
+++ b/src/llama-kv-cells.h
@@ -11,7 +11,7 @@
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
-class llama_kv_cells_unified {
+class llama_kv_cells {
 public:
     void reset() {
         for (uint32_t i = 0; i < pos.size(); ++i) {
@@ -77,30 +77,30 @@ class llama_kv_cells_unified {
     }
 
     // move cell isrc to idst (used during defrag)
-    void mv(uint32_t isrc, uint32_t idst) {
-        assert(isrc < pos.size());
-        assert(idst < pos.size());
+    //void mv(uint32_t isrc, uint32_t idst) {
+    //    assert(isrc < pos.size());
+    //    assert(idst < pos.size());
 
-        assert(pos[idst] == -1);
-        assert(pos[isrc] != -1);
+    //    assert(pos[idst] == -1);
+    //    assert(pos[isrc] != -1);
 
-        pos  [idst] = pos  [isrc];
-        shift[idst] = shift[isrc];
-        seq  [idst] = seq  [isrc];
+    //    pos  [idst] = pos  [isrc];
+    //    shift[idst] = shift[isrc];
+    //    seq  [idst] = seq  [isrc];
 
-        pos  [isrc] = -1;
-        shift[isrc] =  0;
-        seq  [isrc].reset();
+    //    pos  [isrc] = -1;
+    //    shift[isrc] =  0;
+    //    seq  [isrc].reset();
 
-        used.erase (isrc);
-        used.insert(idst);
-    }
+    //    used.erase (isrc);
+    //    used.insert(idst);
+    //}
 
     // copy the state of cells [i, i + n) (used for save/restore the state of the cells)
-    llama_kv_cells_unified cp(uint32_t i, uint32_t n) const {
+    llama_kv_cells cp(uint32_t i, uint32_t n) const {
         assert(i + n <= pos.size());
 
-        llama_kv_cells_unified res;
+        llama_kv_cells res;
 
         res.resize(n);
 
@@ -117,8 +117,8 @@ class llama_kv_cells_unified {
     }
 
     // copy the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
-    llama_kv_cells_unified cp(const std::vector<uint32_t> & idxs) const {
-        llama_kv_cells_unified res;
+    llama_kv_cells cp(const std::vector<uint32_t> & idxs) const {
+        llama_kv_cells res;
 
         res.resize(idxs.size());
 
@@ -135,7 +135,7 @@ class llama_kv_cells_unified {
     }
 
     // set the state of cells [i, i + other.pos.size()) (used for save/restore the state of the cells)
-    void set(uint32_t i, const llama_kv_cells_unified & other) {
+    void set(uint32_t i, const llama_kv_cells & other) {
         assert(i + other.pos.size() <= pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {
@@ -165,7 +165,7 @@ class llama_kv_cells_unified {
     }
 
     // set the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
-    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells_unified & other) {
+    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells & other) {
         assert(idxs.size() == other.pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {
diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp
index e98b4e3546959..ba61ebaa885fe 100644
--- a/src/llama-memory-hybrid.cpp
+++ b/src/llama-memory-hybrid.cpp
@@ -9,32 +9,29 @@
 //
 
 llama_memory_hybrid::llama_memory_hybrid(
-    const llama_model & model,
-                         /* attn */
-            ggml_type    type_k,
-            ggml_type    type_v,
-                 bool    v_trans,
-             uint32_t    kv_size,
-             uint32_t    n_pad,
-             uint32_t    n_swa,
-       llama_swa_type    swa_type,
-                         /* recurrent */
-            ggml_type    type_r,
-            ggml_type    type_s,
-             uint32_t    rs_size,
-                         /* common */
-             uint32_t    n_seq_max,
-                 bool    offload,
-                 bool    unified,
-                         /* layer filters */
-      layer_filter_cb && filter_attn,
-      layer_filter_cb && filter_recr) :
+        const llama_model & model,
+                            /* attn */
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                 uint32_t   kv_size,
+                 uint32_t   n_pad,
+                 uint32_t   n_swa,
+           llama_swa_type   swa_type,
+                            /* recurrent */
+                ggml_type   type_r,
+                ggml_type   type_s,
+                 uint32_t   rs_size,
+                            /* common */
+                 uint32_t   n_seq_max,
+                     bool   offload,
+                     bool   unified,
+                            /* layer filters */
+    const layer_filter_cb & filter_attn,
+    const layer_filter_cb & filter_recr) :
     hparams(model.hparams),
-    mem_attn(new llama_kv_cache_unified(
+    mem_attn(new llama_kv_cache(
         model,
-        filter_attn == nullptr ?
-            [&](int32_t il) { return !hparams.is_recurrent(il); }
-            : filter_attn,
         type_k,
         type_v,
         v_trans,
@@ -44,18 +41,22 @@ llama_memory_hybrid::llama_memory_hybrid(
         n_seq_max,
         n_pad,
         n_swa,
-        swa_type
+        swa_type,
+        filter_attn == nullptr ?
+            [&](int32_t il) { return !hparams.is_recurrent(il); }
+            : filter_attn,
+        nullptr
     )),
     mem_recr(new llama_memory_recurrent(
         model,
-        filter_recr == nullptr ?
-            [&](int32_t il) { return hparams.is_recurrent(il); }
-            : filter_recr,
         type_r,
         type_s,
         offload,
         rs_size,
-        n_seq_max
+        n_seq_max,
+        filter_recr == nullptr ?
+            [&](int32_t il) { return hparams.is_recurrent(il); }
+            : filter_recr
     )) {}
 
 llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
@@ -165,17 +166,21 @@ llama_pos llama_memory_hybrid::seq_pos_max(llama_seq_id seq_id) const {
     return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
 }
 
-void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     mem_attn->state_write(io, seq_id);
     mem_recr->state_write(io, seq_id);
 }
 
-void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     mem_attn->state_read(io, seq_id);
     mem_recr->state_read(io, seq_id);
 }
 
-llama_kv_cache_unified * llama_memory_hybrid::get_mem_attn() const {
+llama_kv_cache * llama_memory_hybrid::get_mem_attn() const {
     return mem_attn.get();
 }
 
@@ -206,7 +211,7 @@ llama_memory_hybrid_context::llama_memory_hybrid_context(
         std::vector<llama_ubatch>   ubatches) :
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
-    ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
+    ctx_attn(new llama_kv_cache_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
     ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(),                        this->ubatches)),
     status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }
@@ -244,8 +249,8 @@ const llama_ubatch & llama_memory_hybrid_context::get_ubatch() const {
     return ubatches[i_next];
 }
 
-const llama_kv_cache_unified_context * llama_memory_hybrid_context::get_attn() const {
-    return static_cast<const llama_kv_cache_unified_context *>(ctx_attn.get());
+const llama_kv_cache_context * llama_memory_hybrid_context::get_attn() const {
+    return static_cast<const llama_kv_cache_context *>(ctx_attn.get());
 }
 
 const llama_memory_recurrent_context * llama_memory_hybrid_context::get_recr() const {
diff --git a/src/llama-memory-hybrid.h b/src/llama-memory-hybrid.h
index c2d56cd541594..11a3565178297 100644
--- a/src/llama-memory-hybrid.h
+++ b/src/llama-memory-hybrid.h
@@ -2,7 +2,7 @@
 
 #include "llama-batch.h"
 #include "llama-graph.h"
-#include "llama-kv-cache-unified.h"
+#include "llama-kv-cache.h"
 #include "llama-memory.h"
 #include "llama-memory-recurrent.h"
 
@@ -13,36 +13,32 @@
 // llama_memory_hybrid
 //
 
-// utilizes instances of llama_memory_recurrent and llama_kv_cache_unified to
+// utilizes instances of llama_memory_recurrent and llama_kv_cache to
 //   support models where each layer may be either attention-based or recurrent
 
 class llama_memory_hybrid : public llama_memory_i {
 public:
-
-    // this callback is used to filter out layers that should not be included in the cache
-    using layer_filter_cb = std::function<bool(int32_t il)>;
-
     llama_memory_hybrid(
         const llama_model & model,
                             /* attn */
-                ggml_type    type_k,
-                ggml_type    type_v,
-                     bool    v_trans,
-                 uint32_t    kv_size,
-                 uint32_t    n_pad,
-                 uint32_t    n_swa,
-           llama_swa_type    swa_type,
-                             /* recurrent */
-                ggml_type    type_r,
-                ggml_type    type_s,
-                 uint32_t    rs_size,
-                             /* common */
-                 uint32_t    n_seq_max,
-                     bool    offload,
-                     bool    unified,
-                             /* layer filters */
-          layer_filter_cb && filter_attn = nullptr,
-          layer_filter_cb && filter_recr = nullptr);
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                 uint32_t   kv_size,
+                 uint32_t   n_pad,
+                 uint32_t   n_swa,
+           llama_swa_type   swa_type,
+                            /* recurrent */
+                ggml_type   type_r,
+                ggml_type   type_s,
+                 uint32_t   rs_size,
+                            /* common */
+                 uint32_t   n_seq_max,
+                     bool   offload,
+                     bool   unified,
+                            /* layer filters */
+    const layer_filter_cb & filter_attn = nullptr,
+    const layer_filter_cb & filter_recr = nullptr);
 
     ~llama_memory_hybrid() = default;
 
@@ -74,26 +70,26 @@ class llama_memory_hybrid : public llama_memory_i {
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0)       override;
 
     //
     // llama_memory_hybrid specific API
     //
 
-    llama_kv_cache_unified * get_mem_attn() const;
+    llama_kv_cache * get_mem_attn() const;
     llama_memory_recurrent * get_mem_recr() const;
 
 private:
     const llama_hparams & hparams;
 
-    const std::unique_ptr<llama_kv_cache_unified> mem_attn;
+    const std::unique_ptr<llama_kv_cache> mem_attn;
     const std::unique_ptr<llama_memory_recurrent> mem_recr;
 };
 
 class llama_memory_hybrid_context : public llama_memory_context_i {
 public:
-    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+    using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
 
     // init failure
     explicit llama_memory_hybrid_context(llama_memory_status status);
@@ -125,7 +121,7 @@ class llama_memory_hybrid_context : public llama_memory_context_i {
     // llama_memory_hybrid_context
     //
 
-    const llama_kv_cache_unified_context * get_attn() const;
+    const llama_kv_cache_context * get_attn() const;
     const llama_memory_recurrent_context * get_recr() const;
 
 private:
diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp
index c0c2ec084dc14..08716ed91aed1 100644
--- a/src/llama-memory-recurrent.cpp
+++ b/src/llama-memory-recurrent.cpp
@@ -16,13 +16,13 @@
 //
 
 llama_memory_recurrent::llama_memory_recurrent(
-        const llama_model &  model,
-          layer_filter_cb && filter,
-                ggml_type    type_r,
-                ggml_type    type_s,
-                     bool    offload,
-                 uint32_t    mem_size,
-                 uint32_t    n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
+        const llama_model & model,
+                ggml_type   type_r,
+                ggml_type   type_s,
+                     bool   offload,
+                 uint32_t   mem_size,
+                 uint32_t   n_seq_max,
+    const layer_filter_cb & filter) : hparams(model.hparams), n_seq_max(n_seq_max) {
     const int32_t n_layer = hparams.n_layer;
 
     head = 0;
@@ -680,7 +680,9 @@ size_t llama_memory_recurrent::size_s_bytes() const {
     return size_s_bytes;
 }
 
-void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     std::vector<std::pair<uint32_t, uint32_t>> cell_ranges; // ranges, from inclusive, to exclusive
     uint32_t cell_count = 0;
 
@@ -718,7 +720,9 @@ void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq
     state_write_data(io, cell_ranges);
 }
 
-void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     uint32_t cell_count;
     io.read_to(&cell_count, sizeof(cell_count));
 
diff --git a/src/llama-memory-recurrent.h b/src/llama-memory-recurrent.h
index 4d094f9a05788..c4daf00495bc2 100644
--- a/src/llama-memory-recurrent.h
+++ b/src/llama-memory-recurrent.h
@@ -12,21 +12,17 @@
 //
 
 // TODO: extract the cache state used for graph computation into llama_memory_recurrent_context_i
-//       see the implementation of llama_kv_cache_unified_context_i for an example how to do it
+//       see the implementation of llama_kv_cache_context_i for an example how to do it
 class llama_memory_recurrent : public llama_memory_i {
 public:
-
-    // this callback is used to filter out layers that should not be included in the cache
-    using layer_filter_cb = std::function<bool(int32_t il)>;
-
     llama_memory_recurrent(
-            const llama_model &  model,
-              layer_filter_cb && filter,
-                    ggml_type    type_r,
-                    ggml_type    type_s,
-                         bool    offload,
-                     uint32_t    mem_size,
-                     uint32_t    n_seq_max);
+            const llama_model & model,
+                    ggml_type   type_r,
+                    ggml_type   type_s,
+                         bool   offload,
+                     uint32_t   mem_size,
+                     uint32_t   n_seq_max,
+        const layer_filter_cb & filter);
 
     ~llama_memory_recurrent() = default;
 
@@ -63,8 +59,8 @@ class llama_memory_recurrent : public llama_memory_i {
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
     uint32_t size = 0; // total number of cells, shared across all sequences
diff --git a/src/llama-memory.h b/src/llama-memory.h
index e8ba336e8525d..ccd1f073b0848 100644
--- a/src/llama-memory.h
+++ b/src/llama-memory.h
@@ -3,6 +3,7 @@
 #include "llama.h"
 
 #include <memory>
+#include <functional>
 
 struct llama_ubatch;
 
@@ -36,8 +37,8 @@ bool llama_memory_status_is_fail(llama_memory_status status);
 
 // the interface for managing the memory context during batch processing
 // this interface is implemented per memory type. see:
-//   - llama_kv_cache_unified_context
-//   - llama_kv_cache_unified_iswa_context
+//   - llama_kv_cache_context
+//   - llama_kv_cache_iswa_context
 //   ...
 //
 // the only method that should mutate the memory and the memory context is llama_memory_i::apply()
@@ -64,6 +65,13 @@ using llama_memory_context_ptr = std::unique_ptr<llama_memory_context_i>;
 // general concept of LLM memory
 // the KV cache is a type of LLM memory, but there can be other types
 struct llama_memory_i {
+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
+    // this callback is used to specify which layers should reuse memory from other layers
+    // return negative value to indicate that the layer il should not reuse memory
+    using layer_reuse_cb = std::function<int32_t(int32_t il)>;
+
     virtual ~llama_memory_i() = default;
 
     // split the input batch into a set of ubatches and verify that they can fit into the cache
@@ -77,7 +85,7 @@ struct llama_memory_i {
     // simulate full cache, used for allocating worst-case compute buffers
     virtual llama_memory_context_ptr init_full() = 0;
 
-    // prepare for any pending memory updates, such as shifts, defrags, etc.
+    // prepare for any pending memory updates, such as shifts, copies, etc.
     // status == LLAMA_MEMORY_STATUS_NO_UPDATE if there is nothing to update
     virtual llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) = 0;
 
@@ -104,13 +112,8 @@ struct llama_memory_i {
     // state write/read
     //
 
-    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
-    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) = 0;
 };
 
 using llama_memory_ptr = std::unique_ptr<llama_memory_i>;
-
-// TODO: temporary until the llama_kv_cache is removed from the public API
-struct llama_kv_cache : public llama_memory_i {
-    virtual ~llama_kv_cache() = default;
-};
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 58ca7df707ef3..7d3429617bef9 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -6,8 +6,8 @@
 #include "llama-cparams.h"
 #include "llama-model-loader.h"
 
-#include "llama-kv-cache-unified.h"
-#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
 #include "llama-memory-recurrent.h"
 
@@ -83,9 +83,11 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_32B:           return "32B";
         case LLM_TYPE_34B:           return "34B";
         case LLM_TYPE_35B:           return "35B";
+        case LLM_TYPE_36B:           return "36B";
         case LLM_TYPE_40B:           return "40B";
         case LLM_TYPE_65B:           return "65B";
         case LLM_TYPE_70B:           return "70B";
+        case LLM_TYPE_120B:          return "120B";
         case LLM_TYPE_142B:          return "142B";
         case LLM_TYPE_236B:          return "236B";
         case LLM_TYPE_290B:          return "290B";
@@ -1095,6 +1097,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_layer) {
+                    case 18: type = LLM_TYPE_537M; break;
                     case 26: type = LLM_TYPE_1B; break;
                     case 34: type = LLM_TYPE_4B; break;
                     case 48: type = LLM_TYPE_12B; break;
@@ -1112,6 +1115,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                 hparams.set_swa_pattern(5);
 
+                hparams.n_layer_kv_from_start     = 20;
                 hparams.rope_freq_base_train_swa  = 10000.0f;
                 hparams.rope_freq_scale_train_swa = 1.0f;
                 hparams.f_attention_scale         = 1.0f;
@@ -1286,6 +1290,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_SEED_OSS:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 64: type = LLM_TYPE_36B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_OLMOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1463,12 +1475,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 // Expert gating function (GLM-4.5 uses sigmoid)
                 ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
                 if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
-                    hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+                    hparams.expert_gating_func =  LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
                 }
 
                 // NextN/MTP parameters
                 ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.nextn_predict_layers, false);
 
+                // TODO: when MTP is implemented, this should probably be updated if needed
+                hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
                 switch (hparams.n_layer) {
                     case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer)
                     case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer)
@@ -1833,7 +1848,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                 hparams.set_swa_pattern(2);
 
-                // TODO: switch (hparams.n_layer)
+                switch (hparams.n_layer) {
+                    case 24: type = LLM_TYPE_20B; break;
+                    case 36: type = LLM_TYPE_120B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
             } break;
         case LLM_ARCH_LFM2:
             {
@@ -3961,6 +3980,43 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_SEED_OSS:
+                {
+                    const uint32_t head_dim             = hparams.n_embd_head_k;
+                    const int64_t n_qo_dim              = n_head * head_dim;
+                    const int64_t n_kv_dim              = n_head_kv * head_dim;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_qo_dim}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_kv_dim}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_kv_dim}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_qo_dim, n_embd}, 0);
+
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_qo_dim},   TENSOR_NOT_REQUIRED);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_kv_dim},   TENSOR_NOT_REQUIRED);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_kv_dim},   TENSOR_NOT_REQUIRED);
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                    }
+                } break;
+
             case LLM_ARCH_OLMOE:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5468,8 +5524,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_LFM2:
                 {
-                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
                     tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
+                    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,          "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -5980,7 +6041,7 @@ struct llm_build_llama : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
@@ -6044,7 +6105,7 @@ struct llm_build_llama : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
 
@@ -6140,7 +6201,7 @@ struct llm_build_llama_iswa : public llm_graph_context {
         ggml_tensor * inp_attn_scale = nullptr;
         inp_attn_scale = build_inp_attn_scale();
 
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
@@ -6218,7 +6279,7 @@ struct llm_build_llama_iswa : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
 
@@ -6319,7 +6380,7 @@ struct llm_build_deci : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
@@ -6395,7 +6456,7 @@ struct llm_build_deci : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6475,7 +6536,7 @@ struct llm_build_baichuan : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -6527,7 +6588,7 @@ struct llm_build_baichuan : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6597,7 +6658,7 @@ struct llm_build_xverse : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -6642,7 +6703,7 @@ struct llm_build_xverse : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6711,7 +6772,7 @@ struct llm_build_falcon : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -6742,9 +6803,9 @@ struct llm_build_falcon : public llm_graph_context {
 
                 ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                 ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 // using mode = 2 for neox mode
                 Qcur = ggml_rope_ext(
@@ -6765,7 +6826,7 @@ struct llm_build_falcon : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6835,7 +6896,7 @@ struct llm_build_grok : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -6895,7 +6956,7 @@ struct llm_build_grok : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -6995,7 +7056,7 @@ struct llm_build_dbrx : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -7022,9 +7083,9 @@ struct llm_build_dbrx : public llm_graph_context {
 
                 Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                 Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
@@ -7044,7 +7105,7 @@ struct llm_build_dbrx : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7119,7 +7180,7 @@ struct llm_build_starcoder : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
         cb(pos, "pos_embd", -1);
@@ -7144,13 +7205,13 @@ struct llm_build_starcoder : public llm_graph_context {
                 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
                 cb(cur, "bqkv", il);
 
-                ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd));
+                ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_cont_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
@@ -7158,7 +7219,7 @@ struct llm_build_starcoder : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7224,7 +7285,7 @@ struct llm_build_refact : public llm_graph_context {
 
         inpL = build_inp_embd(model.tok_embd);
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -7257,7 +7318,7 @@ struct llm_build_refact : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7366,13 +7427,15 @@ struct llm_build_bert : public llm_graph_context {
                         cb(cur, "bqkv", il);
                     }
 
-                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                    Qcur = ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd));
+                    Kcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd));
+                    Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+                    Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 } else {
                     Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq);
                     Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
                     Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 }
 
                 if (model.layers[il].attn_q_norm) {
@@ -7380,6 +7443,10 @@ struct llm_build_bert : public llm_graph_context {
                             model.layers[il].attn_q_norm,
                             model.layers[il].attn_q_norm_b,
                             LLM_NORM, il);
+
+                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                } else {
+                    Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 }
 
                 if (model.layers[il].attn_k_norm) {
@@ -7387,11 +7454,11 @@ struct llm_build_bert : public llm_graph_context {
                             model.layers[il].attn_k_norm,
                             model.layers[il].attn_k_norm_b,
                             LLM_NORM, il);
-                }
 
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                } else {
+                    Kcur = ggml_cont_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                }
 
                 // RoPE
                 if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
@@ -7414,7 +7481,7 @@ struct llm_build_bert : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -7536,9 +7603,9 @@ struct llm_build_neo_bert : public llm_graph_context {
 
                 Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                 Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 // RoPE
                 Qcur = ggml_rope_ext(
@@ -7559,7 +7626,7 @@ struct llm_build_neo_bert : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -7620,7 +7687,7 @@ struct llm_build_bloom : public llm_graph_context {
 
         inpL = build_inp_embd(model.tok_embd);
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         inpL = build_norm(inpL,
                 model.tok_norm,
@@ -7645,13 +7712,13 @@ struct llm_build_bloom : public llm_graph_context {
                 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
                 cb(cur, "bqkv", il);
 
-                ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd));
+                ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_cont_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
@@ -7659,7 +7726,7 @@ struct llm_build_bloom : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7727,7 +7794,7 @@ struct llm_build_mpt : public llm_graph_context {
 
         inpL = build_inp_embd(model.tok_embd);
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         if (model.pos_embd) {
             // inp_pos - contains the positions
@@ -7769,7 +7836,7 @@ struct llm_build_mpt : public llm_graph_context {
 
                 ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd));
                 ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
@@ -7788,17 +7855,18 @@ struct llm_build_mpt : public llm_graph_context {
                             model.layers[il].attn_k_norm_b,
                             LLM_NORM, il);
                     cb(Kcur, "Kcur", il);
+
+                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
                 } else {
-                    Qcur = ggml_cont(ctx0, Qcur);
+                    Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                     cb(Qcur, "Qcur", il);
 
-                    Kcur = ggml_cont(ctx0, Kcur);
+                    Kcur = ggml_cont_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
                     cb(Kcur, "Kcur", il);
                 }
 
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
@@ -7806,7 +7874,7 @@ struct llm_build_mpt : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -7876,7 +7944,7 @@ struct llm_build_stablelm : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -7952,7 +8020,7 @@ struct llm_build_stablelm : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8028,7 +8096,7 @@ struct llm_build_qwen : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -8050,9 +8118,9 @@ struct llm_build_qwen : public llm_graph_context {
 
                 ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,   n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                 ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd)));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd));
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 // using mode = 2 for neox mode
                 Qcur = ggml_rope_ext(
@@ -8073,7 +8141,7 @@ struct llm_build_qwen : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8143,7 +8211,7 @@ struct llm_build_qwen2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -8193,7 +8261,7 @@ struct llm_build_qwen2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8307,8 +8375,9 @@ struct llm_build_dream : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                cur = build_attn(inp_attn, model.layers[il].wo, model.layers[il].bo, Qcur, Kcur, Vcur, nullptr,
-                                 nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8407,8 +8476,9 @@ struct llm_build_llada : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                cur = build_attn(inp_attn, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr,
-                                 1.0f / sqrtf(float(n_embd_head)), il);
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8468,7 +8538,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         int sections[4];
         std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
@@ -8521,7 +8591,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8589,7 +8659,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -8648,7 +8718,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8748,7 +8818,7 @@ struct llm_build_qwen3 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -8801,7 +8871,7 @@ struct llm_build_qwen3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8869,7 +8939,7 @@ struct llm_build_qwen3moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -8922,7 +8992,7 @@ struct llm_build_qwen3moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -8999,7 +9069,7 @@ struct llm_build_phi2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -9025,21 +9095,21 @@ struct llm_build_phi2 : public llm_graph_context {
 
                     Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                     Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                    Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+                    Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 } else {
                     Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq);
                     Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk);
                     Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv);
                     Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                     Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 }
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -9062,7 +9132,7 @@ struct llm_build_phi2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9128,13 +9198,13 @@ struct llm_build_phi3 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
+        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
         inp_attn_type * inp_attn = nullptr;
 
         if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_unified_iswa();
+            inp_attn = build_attn_inp_kv_iswa();
         } else {
-            inp_attn = build_attn_inp_kv_unified();
+            inp_attn = build_attn_inp_kv();
         }
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -9163,21 +9233,21 @@ struct llm_build_phi3 : public llm_graph_context {
 
                     Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head * sizeof(float), cur->nb[1], 0 * sizeof(float) * (n_embd));
                     Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)));
+                    Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
+                    Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 } else {
                     Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq);
                     Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk);
                     Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv);
                     Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                     Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 }
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, rope_factors,
                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -9199,7 +9269,7 @@ struct llm_build_phi3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9286,7 +9356,7 @@ struct llm_build_plamo : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -9333,7 +9403,7 @@ struct llm_build_plamo : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9402,7 +9472,7 @@ struct llm_build_gpt2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
         cb(pos, "pos_embd", -1);
@@ -9427,21 +9497,21 @@ struct llm_build_gpt2 : public llm_graph_context {
                 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
                 cb(cur, "bqkv", il);
 
-                ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd));
+                ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_cont_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9512,7 +9582,7 @@ struct llm_build_codeshell : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -9533,9 +9603,9 @@ struct llm_build_codeshell : public llm_graph_context {
 
                 ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                 ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
@@ -9555,7 +9625,7 @@ struct llm_build_codeshell : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9625,7 +9695,7 @@ struct llm_build_orion : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -9684,7 +9754,7 @@ struct llm_build_orion : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9752,7 +9822,7 @@ struct llm_build_internlm2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -9811,7 +9881,7 @@ struct llm_build_internlm2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -9888,7 +9958,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -9999,7 +10069,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -10083,7 +10153,7 @@ struct llm_build_gemma : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -10129,7 +10199,7 @@ struct llm_build_gemma : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -10199,7 +10269,7 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -10244,7 +10314,7 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -10333,7 +10403,7 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
         ggml_tensor * inp_pos = build_inp_pos();
 
         // TODO: is causal == true correct? might need some changes
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -10386,7 +10456,7 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -10458,7 +10528,6 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
     const int64_t n_embd_altup;
     const int64_t n_altup;
     const int     i_altup_act;
-    const int     n_layer_kv = 20; // number of layers having KV [KV_REUSE]
     const int     n_layer_sparsity = 10; // number of layers using activation sparsity
     const float   f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95)
 
@@ -10484,7 +10553,7 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
         ggml_tensor * inp_pos = build_inp_pos();
 
         // TODO: is causal == true correct? might need some changes
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
 
         // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer]
         ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
@@ -10508,8 +10577,6 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
 
         for (int il = 0; il < n_layer; ++il) {
             // this block is made to be closely resemble Gemma3p5DecoderLayer on python code
-            const bool has_kv = (il < n_layer_kv);
-
             const float freq_base_l  = model.get_rope_freq_base (cparams, il);
             const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
 
@@ -10529,7 +10596,7 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
             ggml_tensor * laurel_out = laurel(cur, il); // [n_embd, n_tokens]
 
             // self-attention
-            if (has_kv) {
+            if (hparams.has_kv(il)) {
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
@@ -10567,9 +10634,9 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
             } else {
-                // no KV layers
+                // reuse KV cache of earlier layers
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
                 cb(Qcur, "Qcur", il);
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
@@ -10585,7 +10652,7 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                     model.layers[il].wo, NULL,
-                    Qcur, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
+                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
             }
 
             cur = build_norm(cur,
@@ -10863,8 +10930,8 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
         ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities); // [n_altup, n_tokens]
         all_coefs = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f); // + 1.0
         cb(all_coefs, "all_coefs", il);
-        all_coefs = ggml_cont(ctx0, ggml_transpose(ctx0, all_coefs)); // [n_tokens, n_altup]
-        all_coefs = ggml_reshape_3d(ctx0, all_coefs, 1, n_tokens, n_altup); // [1, n_tokens, n_altup]
+        all_coefs = ggml_transpose(ctx0, all_coefs); // [n_tokens, n_altup]
+        all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup); // [1, n_tokens, n_altup]
 
         innovation = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1);
         ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs); // [n_embd, n_tokens, n_altup]
@@ -10891,7 +10958,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -10950,7 +11017,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11377,7 +11444,9 @@ struct llm_build_jamba : public llm_graph_context_mamba {
                 cb(Vcur, "Vcur", il);
 
                 // No RoPE :)
-                cur = build_attn(inp_hybrid->get_attn(), model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il);
+                cur = build_attn(inp_hybrid->get_attn(),
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11460,7 +11529,7 @@ struct llm_build_command_r : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -11535,7 +11604,7 @@ struct llm_build_command_r : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11607,7 +11676,7 @@ struct llm_build_cohere2_iswa : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -11670,7 +11739,7 @@ struct llm_build_cohere2_iswa : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11742,7 +11811,7 @@ struct llm_build_olmo : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -11801,7 +11870,7 @@ struct llm_build_olmo : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11870,7 +11939,7 @@ struct llm_build_olmo2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -11921,7 +11990,7 @@ struct llm_build_olmo2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -11999,7 +12068,7 @@ struct llm_build_olmoe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -12054,7 +12123,7 @@ struct llm_build_olmoe : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12125,7 +12194,7 @@ struct llm_build_openelm : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -12187,7 +12256,7 @@ struct llm_build_openelm : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12256,7 +12325,7 @@ struct llm_build_gptneox : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -12277,9 +12346,9 @@ struct llm_build_gptneox : public llm_graph_context {
 
                 ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                 ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
@@ -12299,7 +12368,7 @@ struct llm_build_gptneox : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12402,7 +12471,7 @@ struct llm_build_arctic : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -12449,7 +12518,7 @@ struct llm_build_arctic : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12540,7 +12609,7 @@ struct llm_build_deepseek : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
@@ -12604,7 +12673,7 @@ struct llm_build_deepseek : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -12717,7 +12786,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -12832,7 +12901,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
                     // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
                     cur = build_attn(inp_attn,
                             model.layers[il].wo, NULL,
-                            Qcur, Kcur, Vcur, nullptr, model.layers[il].wv_b, kq_scale, il);
+                            Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
                 } else {
                     ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
                     cb(kv, "kv", il);
@@ -12866,7 +12935,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
                     // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
                     cur = build_attn(inp_attn,
                             model.layers[il].wo, NULL,
-                            Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 }
             }
 
@@ -12964,7 +13033,7 @@ struct llm_build_bitnet : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -13033,7 +13102,7 @@ struct llm_build_bitnet : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         NULL, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 
                 cur = build_norm(cur,
                         model.layers[il].attn_sub_norm, NULL,
@@ -13156,7 +13225,7 @@ struct llm_build_t5_enc : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo_enc, nullptr,
-                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -13228,7 +13297,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
         const int64_t n_outputs_enc = embd_enc->ne[1];
 
-        auto * inp_attn_self  = build_attn_inp_kv_unified();
+        auto * inp_attn_self  = build_attn_inp_kv();
         auto * inp_attn_cross = build_attn_inp_cross();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -13262,7 +13331,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
                 cur = build_attn(inp_attn_self,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -13294,7 +13363,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
                 cur = build_attn(inp_attn_cross,
                         model.layers[il].wo_cross, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
 
                 //ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
@@ -13393,7 +13462,7 @@ struct llm_build_jais : public llm_graph_context {
 
         inpL = build_inp_embd(model.tok_embd);
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -13412,21 +13481,21 @@ struct llm_build_jais : public llm_graph_context {
                 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
                 cb(cur, "bqkv", il);
 
-                ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd)));
-                ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd)));
-                ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa)));
+                ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd));
+                ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd));
+                ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa));
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+                Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_cont_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -13491,7 +13560,7 @@ struct llm_build_chatglm : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -13525,6 +13594,7 @@ struct llm_build_chatglm : public llm_graph_context {
                     }
                     Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                     Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 } else {
                     cur = build_lora_mm(model.layers[il].wqkv, cur);
                     cb(cur, "wqkv", il);
@@ -13534,11 +13604,10 @@ struct llm_build_chatglm : public llm_graph_context {
                     }
                     Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                     Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                    Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+                    Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 }
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
                 //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
@@ -13558,7 +13627,7 @@ struct llm_build_chatglm : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -13624,7 +13693,7 @@ struct llm_build_glm4 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -13659,6 +13728,7 @@ struct llm_build_glm4 : public llm_graph_context {
                     }
                     Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                     Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 } else {
                     cur = build_lora_mm(model.layers[il].wqkv, cur);
                     cb(cur, "wqkv", il);
@@ -13668,11 +13738,10 @@ struct llm_build_glm4 : public llm_graph_context {
                     }
                     Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
                     Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                    Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
+                    Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
                 }
 
-                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
-
                 Qcur = ggml_rope_ext(
                         ctx0, Qcur, inp_pos, nullptr,
                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -13691,7 +13760,7 @@ struct llm_build_glm4 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -13774,7 +13843,7 @@ struct llm_build_glm4_moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -13840,7 +13909,7 @@ struct llm_build_glm4_moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_transformer_layers - 1 && inp_out_ids) {
@@ -13934,7 +14003,7 @@ struct llm_build_nemotron : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -13994,7 +14063,7 @@ struct llm_build_nemotron : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -14063,7 +14132,7 @@ struct llm_build_exaone : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -14125,7 +14194,7 @@ struct llm_build_exaone : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -14195,13 +14264,13 @@ struct llm_build_exaone4 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
+        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
         inp_attn_type * inp_attn = nullptr;
 
         if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_unified_iswa();
+            inp_attn = build_attn_inp_kv_iswa();
         } else {
-            inp_attn = build_attn_inp_kv_unified();
+            inp_attn = build_attn_inp_kv();
         }
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -14256,7 +14325,7 @@ struct llm_build_exaone4 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cb(cur, "attn_out", il);
             }
 
@@ -15084,7 +15153,7 @@ struct llm_build_granite : public llm_graph_context {
             inp_pos = build_inp_pos();
         }
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -15135,12 +15204,12 @@ struct llm_build_granite : public llm_graph_context {
     }
 
     ggml_tensor * build_attention_layer(
-              ggml_tensor                     * cur,
-              ggml_tensor                     * inp_pos,
-              llm_graph_input_attn_kv_unified * inp_attn,
-        const llama_model                     & model,
-        const int64_t                           n_embd_head,
-        const int                               il) {
+              ggml_tensor             * cur,
+              ggml_tensor             * inp_pos,
+              llm_graph_input_attn_kv * inp_attn,
+        const llama_model             & model,
+        const int64_t                 n_embd_head,
+        const int                     il) {
 
         // compute Q and K and (optionally) RoPE them
         ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -15191,7 +15260,7 @@ struct llm_build_granite : public llm_graph_context {
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
         cur = build_attn(inp_attn,
                 model.layers[il].wo, model.layers[il].bo,
-                Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
         return cur;
     }
@@ -15354,12 +15423,12 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
     }
 
     ggml_tensor * build_attention_layer(
-              ggml_tensor                     * cur,
-              ggml_tensor                     * inp_pos,
-              llm_graph_input_attn_kv_unified * inp_attn,
-        const llama_model                     & model,
-        const int64_t                           n_embd_head,
-        const int                               il) {
+              ggml_tensor             * cur,
+              ggml_tensor             * inp_pos,
+              llm_graph_input_attn_kv * inp_attn,
+        const llama_model             & model,
+        const int64_t                 n_embd_head,
+        const int                     il) {
 
         // compute Q and K and (optionally) RoPE them
         ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
@@ -15410,7 +15479,7 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
         cur = build_attn(inp_attn,
                 model.layers[il].wo, model.layers[il].bo,
-                Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
         return cur;
     }
@@ -15516,7 +15585,7 @@ struct llm_build_chameleon : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -15595,7 +15664,7 @@ struct llm_build_chameleon : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -15847,7 +15916,7 @@ struct llm_build_plm : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -15951,7 +16020,7 @@ struct llm_build_plm : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -16012,7 +16081,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -16074,7 +16143,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -16161,7 +16230,7 @@ struct llm_build_dots1 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -16214,7 +16283,7 @@ struct llm_build_dots1 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -16311,7 +16380,7 @@ struct llm_build_ernie4_5 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
@@ -16369,7 +16438,7 @@ struct llm_build_ernie4_5 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -16441,7 +16510,7 @@ struct llm_build_ernie4_5_moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -16502,7 +16571,7 @@ struct llm_build_ernie4_5_moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
                 cb(cur, "attn_out", il);
             }
 
@@ -16655,7 +16724,7 @@ struct llm_build_falcon_h1 : public llm_graph_context_mamba {
 
             ggml_tensor * attn_out = build_attn(inp->get_attn(),
                     model.layers[il].wo, NULL,
-                    Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             cb(attn_out, "attn_out", il);
 
             cur = build_norm(inpL,
@@ -16815,7 +16884,7 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
 
 private:
     ggml_tensor * build_plamo2_attn_layer(
-            llm_graph_input_attn_kv_unified * inp,
+            llm_graph_input_attn_kv * inp,
             ggml_tensor * inp_pos,
             ggml_tensor * cur,
             const llama_model & model,
@@ -16839,13 +16908,13 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
 
             ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head, n_tokens, n_embd_head_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv));
             ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv, n_tokens, n_embd_head_k * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv));
-            ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, n_embd_head_v * n_head_kv, n_tokens, qkv->nb[1], v_offset * ggml_element_size(qkv)));
+            ggml_tensor * Vcur = ggml_view_2d(ctx0, qkv, n_embd_head_v * n_head_kv, n_tokens, qkv->nb[1], v_offset * ggml_element_size(qkv));
 
             cb(Qcur, "Qcur", il);
             cb(Kcur, "Kcur", il);
             cb(Vcur, "Vcur", il);
 
-            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv, n_tokens);
+            Vcur = ggml_cont_3d(ctx0, Vcur, n_embd_head_v, n_head_kv, n_tokens);
 
             Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
             cb(Qcur, "Qcur_normed", il);
@@ -16865,7 +16934,9 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
                     ext_factor, attn_factor, beta_fast, beta_slow
                     );
 
-            cur = build_attn(inp, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f/sqrtf(float(n_embd_head_v)), il);
+            cur = build_attn(inp,
+                    model.layers[il].wo, NULL,
+                    Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head_v)), il);
         }
 
         cb(cur, "attn_out", il);
@@ -16912,15 +16983,13 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
         cb(zx, "mamba_in_proj", il);
         // {8192, 5, 1, 1} -> {8192, 1, 5, 1}
         zx = ggml_permute(ctx0, zx, 0, 2, 1, 3);
-        zx = ggml_cont(ctx0, zx);
-        zx = ggml_reshape_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs);
+        zx = ggml_cont_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs);
         cb(zx, "mamba_in_proj_out", il);
 
         // split into z and x
         // => {head_dim * n_heads, n_seq_tokens, n_seqs}
         ggml_tensor * x = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], head_dim*ggml_element_size(zx));
-        x = ggml_cont(ctx0, x);
-        x = ggml_reshape_3d(ctx0, x, head_dim * n_heads, n_seq_tokens, n_seqs);
+        x = ggml_cont_3d(ctx0, x, head_dim * n_heads, n_seq_tokens, n_seqs);
         // x = ggml_permute(ctx0, x, 0, 2, 1, 3);
         cb(x, "mamba_x_split", il);
 
@@ -17050,7 +17119,7 @@ struct llm_build_arcee : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
@@ -17114,7 +17183,7 @@ struct llm_build_arcee : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
 
@@ -17185,7 +17254,7 @@ struct llm_build_hunyuan_moe : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
 
@@ -17259,7 +17328,7 @@ struct llm_build_hunyuan_moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
 
@@ -17346,7 +17415,7 @@ struct llm_build_hunyuan_dense : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
 
@@ -17419,7 +17488,7 @@ struct llm_build_hunyuan_dense : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
 
@@ -17484,7 +17553,7 @@ struct llm_build_smollm3 : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv();
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
@@ -17549,7 +17618,7 @@ struct llm_build_smollm3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
 
@@ -17616,7 +17685,7 @@ struct llm_build_openai_moe_iswa : public llm_graph_context {
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        auto * inp_attn = build_attn_inp_kv_unified_iswa();
+        auto * inp_attn = build_attn_inp_kv_iswa();
 
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
@@ -17671,9 +17740,9 @@ struct llm_build_openai_moe_iswa : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                cur = build_attn_with_sinks(inp_attn,
+                cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].attn_sinks, 1.0f/sqrtf(float(n_rot)), il);
+                        Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il);
 
                 cb(cur, "attn_out", il);
             }
@@ -17770,8 +17839,7 @@ struct llm_build_lfm2 : public llm_graph_context {
         cb(cur, "model.embedding_norm", -1);
         res->t_embd = cur;
 
-        // lm_head is tied with embeddings
-        cur = build_lora_mm(model.tok_embd, cur);
+        cur = build_lora_mm(model.output, cur);
         cb(cur, "lm_head", -1);
 
         res->t_logits = cur;
@@ -17798,10 +17866,10 @@ struct llm_build_lfm2 : public llm_graph_context {
         return cur;
     }
 
-    ggml_tensor * build_attn_block(ggml_tensor                     * cur,
-                                   ggml_tensor                     * inp_pos,
-                                   llm_graph_input_attn_kv_unified * inp_attn,
-                                   int                               il) const {
+    ggml_tensor * build_attn_block(ggml_tensor             * cur,
+                                   ggml_tensor             * inp_pos,
+                                   llm_graph_input_attn_kv * inp_attn,
+                                   int                     il) const {
         GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
         auto const n_embd_head = hparams.n_embd_head_v;
         auto const n_head_kv = hparams.n_head_kv(il);
@@ -17836,7 +17904,7 @@ struct llm_build_lfm2 : public llm_graph_context {
                 );
 
         cur = build_attn(inp_attn, model.layers[il].wo, NULL,
-                q, k, v, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                q, k, v, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 
         cb(cur, "model.layers.{}.self_attn.out_proj", il);
 
@@ -17913,6 +17981,137 @@ struct llm_build_lfm2 : public llm_graph_context {
     }
 };
 
+struct llm_build_seed_oss : public llm_graph_context {
+    llm_build_seed_oss(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
+
+        auto * inp_attn = build_attn_inp_kv();
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                        ctx0, Qcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                Kcur = ggml_rope_ext(
+                        ctx0, Kcur, inp_pos, nullptr,
+                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                        );
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = build_norm(ffn_inp,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_post_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 template <bool iswa>
 struct llm_build_smallthinker : public llm_graph_context{
     llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
@@ -17929,13 +18128,13 @@ struct llm_build_smallthinker : public llm_graph_context{
         // inp_pos - contains the positions
         ggml_tensor * inp_pos = build_inp_pos();
 
-        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_unified_iswa, llm_graph_input_attn_kv_unified>;
+        using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
         inp_attn_type * inp_attn = nullptr;
 
         if constexpr (iswa) {
-            inp_attn = build_attn_inp_kv_unified_iswa();
+            inp_attn = build_attn_inp_kv_iswa();
         } else {
-            inp_attn = build_attn_inp_kv_unified();
+            inp_attn = build_attn_inp_kv();
         }
 
         ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -17980,7 +18179,7 @@ struct llm_build_smallthinker : public llm_graph_context{
 
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1 && inp_out_ids) {
@@ -18058,14 +18257,14 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                 if (llm_arch_is_recurrent(arch)) {
                     res = new llama_memory_recurrent(
                             *this,
-                            nullptr,
                             GGML_TYPE_F32,
                             GGML_TYPE_F32,
                             cparams.offload_kqv,
                             std::max((uint32_t) 1, cparams.n_seq_max),
-                            cparams.n_seq_max);
+                            cparams.n_seq_max,
+                            nullptr);
                 } else if (llm_arch_is_hybrid(arch)) {
-                    const auto padding = llama_kv_cache_unified::get_padding(cparams);
+                    const auto padding = llama_kv_cache::get_padding(cparams);
 
                     cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
 
@@ -18087,7 +18286,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         /* filter_attn       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr,
                         /* filter_recr       */ (arch == LLM_ARCH_FALCON_H1) ? [&](int32_t) { return true; } : (llama_memory_hybrid::layer_filter_cb)nullptr);
                 } else {
-                    const auto padding = llama_kv_cache_unified::get_padding(cparams);
+                    const auto padding = llama_kv_cache::get_padding(cparams);
 
                     uint32_t n_ctx_per_stream = cparams.n_ctx;
 
@@ -18104,10 +18303,22 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
 
                     LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
 
+                    llama_memory_i::layer_reuse_cb reuse = nullptr;
+
+                    if (arch == LLM_ARCH_GEMMA3N) {
+                        reuse = [&](int32_t il) {
+                            if (il >= (int32_t) hparams.n_layer_kv_from_start) {
+                                return (int32_t) hparams.n_layer_kv_from_start - (hparams.is_swa(il) ? 2 : 1);
+                            }
+
+                            return -1;
+                        };
+                    }
+
                     if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
                         GGML_ASSERT(hparams.is_swa_any());
 
-                        res = new llama_kv_cache_unified_iswa(
+                        res = new llama_kv_cache_iswa(
                                 *this,
                                 params.type_k,
                                 params.type_v,
@@ -18118,13 +18329,14 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 n_ctx_per_stream,
                                 cparams.n_seq_max,
                                 cparams.n_ubatch,
-                                padding);
+                                padding,
+                                nullptr,
+                                reuse);
                     } else {
                         GGML_ASSERT(!hparams.is_swa_any());
 
-                        res = new llama_kv_cache_unified(
+                        res = new llama_kv_cache(
                                 *this,
-                                nullptr,
                                 params.type_k,
                                 params.type_v,
                                 !cparams.flash_attn,
@@ -18134,7 +18346,9 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                                 cparams.n_seq_max,
                                 padding,
                                 hparams.n_swa,
-                                hparams.swa_type);
+                                hparams.swa_type,
+                                nullptr,
+                                nullptr);
                     }
                 }
             }
@@ -18451,6 +18665,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_bailingmoe>(*this, params);
             } break;
+        case LLM_ARCH_SEED_OSS:
+            {
+                llm = std::make_unique<llm_build_seed_oss>(*this, params);
+            } break;
         case LLM_ARCH_DOTS1:
             {
                 llm = std::make_unique<llm_build_dots1>(*this, params);
@@ -18509,6 +18727,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     return llm->res->get_gf();
 }
 
+
 //
 // interface implementation
 //
@@ -18703,6 +18922,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_SMALLTHINKER:
         case LLM_ARCH_GLM4_MOE:
+        case LLM_ARCH_SEED_OSS:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:
diff --git a/src/llama-model.h b/src/llama-model.h
index 6fcd74d57fdca..af4460cc01eb0 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -39,6 +39,7 @@ enum llm_type {
     LLM_TYPE_410M,
     LLM_TYPE_450M,
     LLM_TYPE_475M,
+    LLM_TYPE_537M,
     LLM_TYPE_700M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,
@@ -75,9 +76,11 @@ enum llm_type {
     LLM_TYPE_32B,
     LLM_TYPE_34B,
     LLM_TYPE_35B,
+    LLM_TYPE_36B,
     LLM_TYPE_40B,
     LLM_TYPE_65B,
     LLM_TYPE_70B,
+    LLM_TYPE_120B,
     LLM_TYPE_142B,
     LLM_TYPE_236B,
     LLM_TYPE_290B,
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index f7e03e702ed19..de5d1681dff85 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -2341,7 +2341,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
         // @ngxson : quick hack for gpt-oss, always render these tokens
         for (const auto & t : token_to_id) {
-            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>") {
+            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>" || t.first == "<|constrain|>") {
                 id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
             }
         }
@@ -2388,6 +2388,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
             if (has_return && has_call && has_end) {
                 special_eog_ids.erase(end_id);
+                id_to_token[end_id].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
                 LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
             }
         }
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 9658abf969dd2..91719577564a9 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -192,7 +192,10 @@ if (NOT WIN32)
     llama_build_and_test(test-arg-parser.cpp)
 endif()
 
-# llama_build_and_test(test-opt.cpp) # SLOW
+if (NOT LLAMA_SANITIZE_ADDRESS)
+  # TODO: repair known memory leaks
+  llama_build_and_test(test-opt.cpp)
+endif()
 llama_build_and_test(test-gguf.cpp)
 llama_build_and_test(test-backend-ops.cpp)
 
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index d29779cd12b22..74886b4549056 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -2491,12 +2491,12 @@ struct test_bin_bcast : public test_case {
         : op(op), type(type), ne(ne), nr(nr), nf(nf) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        GGML_ASSERT(nf <= 8);
+        GGML_ASSERT(nf <= 16);
 
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0]*nr[0], ne[1]*nr[1], ne[2]*nr[2], ne[3]*nr[3]);
         ggml_set_name(a, "a");
 
-        ggml_tensor * b[8];
+        ggml_tensor * b[16];
         for (int i = 0; i < nf; ++i) {
             b[i] = ggml_new_tensor(ctx, type, 4, ne.data());
             ggml_set_name(b[i], (std::string("b") + std::to_string(i)).c_str());
@@ -2858,6 +2858,7 @@ struct test_rms_norm_mul_add : public test_case {
     const std::array<int64_t, 4> ne;
     const float eps;
     const bool broadcast;
+    const bool multi_add; // test a sequence of adds feeding into rms_norm
 
     std::string op_desc(ggml_tensor * t) override {
         GGML_UNUSED(t);
@@ -2867,13 +2868,13 @@ struct test_rms_norm_mul_add : public test_case {
     bool run_whole_graph() override { return true; }
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, eps, broadcast);
+        return VARS_TO_STR5(type, ne, eps, broadcast, multi_add);
     }
 
     test_rms_norm_mul_add(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 5, 4, 3},
-            float eps = 1e-6f, bool broadcast = false)
-        : type(type), ne(ne), eps(eps), broadcast(broadcast) {}
+            float eps = 1e-6f, bool broadcast = false, bool multi_add = false)
+        : type(type), ne(ne), eps(eps), broadcast(broadcast), multi_add(multi_add) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         std::array<int64_t, 4> broadcast_dims = {ne[0]*2, ne[1]*3, ne[2]*3, ne[3]*4};
@@ -2891,6 +2892,9 @@ struct test_rms_norm_mul_add : public test_case {
 
         // Use a, b and c early, so we don't end up with an OP_NONE between rms_norm and mul
         a = ggml_add(ctx, ggml_add(ctx, a, b), c);
+        if (multi_add) {
+            a = ggml_add(ctx, ggml_add(ctx, a, b), c);
+        }
         ggml_tensor * out = ggml_add(ctx, ggml_mul(ctx, ggml_rms_norm(ctx, a, eps), b), c);
         ggml_set_name(out, "out");
 
@@ -3098,9 +3102,10 @@ struct test_mul_mat : public test_case {
     const std::array<int64_t, 2> nr;  // repeat in dims 3 and 4
     const std::array<int64_t, 4> per; // permutation of dimensions
     const bool v; // whether a and b are non-contiguous views
+    const uint32_t o; // number of outputs
 
     std::string vars() override {
-        return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v);
+        return VARS_TO_STR10(type_a, type_b, m, n, k, bs, nr, per, v, o);
     }
 
     double max_nmse_err() override {
@@ -3121,8 +3126,8 @@ struct test_mul_mat : public test_case {
             std::array<int64_t, 2> bs = {10, 10},
             std::array<int64_t, 2> nr = {2, 2},
             std::array<int64_t, 4> per = {0, 1, 2, 3},
-            bool v = false)
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v) {}
+            bool v = false, uint32_t o = 1)
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v), o(o) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -3186,9 +3191,21 @@ struct test_mul_mat : public test_case {
 
         ggml_tensor * out = ggml_mul_mat(ctx, a, b);
         ggml_set_name(out, "out");
+        for (uint32_t i = 1; i < o; ++i) {
+            ggml_tensor * out2 = ggml_mul_mat(ctx, a, b);
+            ggml_set_name(out2, "out2");
+            out = ggml_add(ctx, out, out2);
+        }
 
         return out;
     }
+
+    bool run_whole_graph() override { return o > 1; }
+
+    std::string op_desc(ggml_tensor * t) override {
+        GGML_UNUSED(t);
+        return ggml_op_name(GGML_OP_MUL_MAT);
+    }
 };
 
 // GGML_OP_MUL_MAT_ID
@@ -3201,9 +3218,10 @@ struct test_mul_mat_id : public test_case {
     const int64_t m;
     const int64_t n;
     const int64_t k;
+    const uint32_t o; // number of outputs
 
     std::string vars() override {
-        return VARS_TO_STR8(type_a, type_b, n_mats, n_used, b, m, n, k);
+        return VARS_TO_STR9(type_a, type_b, n_mats, n_used, b, m, n, k, o);
     }
 
     double max_nmse_err() override {
@@ -3217,9 +3235,9 @@ struct test_mul_mat_id : public test_case {
 
     test_mul_mat_id(ggml_type type_a = GGML_TYPE_F32, ggml_type type_b = GGML_TYPE_F32,
             int n_mats = 8, int n_used = 2, bool b = false,
-            int64_t m = 32, int64_t n = 32, int64_t k = 32)
+            int64_t m = 32, int64_t n = 32, int64_t k = 32, uint32_t o = 1)
         : type_a(type_a), type_b(type_b), n_mats(n_mats), n_used(n_used), b(b),
-            m(m), n(n), k(k) {
+            m(m), n(n), k(k), o(o) {
             GGML_ASSERT(n_used <= n_mats);
         }
 
@@ -3241,6 +3259,13 @@ struct test_mul_mat_id : public test_case {
         ggml_tensor * out = ggml_mul_mat_id(ctx, as, b, ids);
         ggml_set_name(out, "out");
 
+        for (uint32_t i = 1; i < o; ++i) {
+            ggml_tensor * a2 = ggml_new_tensor_3d(ctx, type_a, k, m, n_mats);
+            ggml_tensor * out2 = ggml_mul_mat_id(ctx, a2, b, ids);
+            ggml_set_name(out2, "out2");
+            out = ggml_add(ctx, out, out2);
+        }
+
         return out;
     }
 
@@ -3264,6 +3289,13 @@ struct test_mul_mat_id : public test_case {
             }
         }
     }
+
+    bool run_whole_graph() override { return o > 1; }
+
+    std::string op_desc(ggml_tensor * t) override {
+        GGML_UNUSED(t);
+        return ggml_op_name(GGML_OP_MUL_MAT_ID);
+    }
 };
 
 // GGML_OP_OUT_PROD
@@ -4063,6 +4095,75 @@ struct test_conv_2d_dw : public test_case {
     }
 };
 
+// GGML_OP_CONV_3D
+struct test_conv_3d : public test_case {
+    // Logical 5D dimensions
+    const int64_t N, IC, ID, IH, IW;
+    const int64_t OC, KD, KH, KW;
+    // Conv params
+    const int s0, s1, s2;
+    const int p0, p1, p2;
+    const int d0, d1, d2;
+    // Types
+    const ggml_type type_kernel;
+
+    std::string op_desc(ggml_tensor * t) override {
+        GGML_UNUSED(t);
+        return "CONV_3D";
+    }
+
+    std::string vars() override {
+        return VARS_TO_STR11(N, IC, ID, IH, IW, OC, KD, KH, KW, s0, s1) + "," +
+               VARS_TO_STR8(s2, p0, p1, p2, d0, d1, d2, type_kernel);
+    }
+
+    double max_nmse_err() override {
+        return 5e-4;
+    }
+
+    uint64_t op_flops(ggml_tensor * t) override {
+        GGML_UNUSED(t);
+        auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
+            return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
+        };
+        const int64_t OD = calc_conv_output_size(ID, KD, s2, p2, d2);
+        const int64_t OH = calc_conv_output_size(IH, KH, s1, p1, d1);
+        const int64_t OW = calc_conv_output_size(IW, KW, s0, p0, d0);
+
+        return (uint64_t)N * OC * OD * OH * OW * (2 * IC * KD * KH * KW - 1);
+    }
+
+    test_conv_3d(
+        int64_t N, int64_t IC, int64_t ID, int64_t IH, int64_t IW,
+        int64_t OC, int64_t KD, int64_t KH, int64_t KW,
+        int s0, int s1, int s2,
+        int p0, int p1, int p2,
+        int d0, int d1, int d2,
+        ggml_type type_kernel
+    ) : N(N), IC(IC), ID(ID), IH(IH), IW(IW),
+        OC(OC), KD(KD), KH(KH), KW(KW),
+        s0(s0), s1(s1), s2(s2),
+        p0(p0), p1(p1), p2(p2),
+        d0(d0), d1(d1), d2(d2),
+        type_kernel(type_kernel) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        // GGML input tensor is packed as [W, H, D, C*N]
+        const int64_t ne_input[] = {IW, IH, ID, IC * N};
+        ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input);
+        ggml_set_name(input, "input");
+
+        // GGML kernel tensor is packed as [KW, KH, KD, IC*OC]
+        const int64_t ne_kernel[] = {KW, KH, KD, IC * OC};
+        ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel);
+        ggml_set_name(kernel, "kernel");
+
+        ggml_tensor * out = ggml_conv_3d(ctx, kernel, input, s0, s1, s2, p0, p1, p2, d0, d1, d2, (int)IC, (int)N, (int)OC);
+        ggml_set_name(out, "out");
+        return out;
+    }
+};
+
 // GGML_OP_CONCAT
 struct test_concat : public test_case {
     const ggml_type type;
@@ -4203,20 +4304,32 @@ struct test_sum : public test_case {
 struct test_sum_rows : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
+    const bool permute;
+    const bool slice;
 
     std::string vars() override {
-        return VARS_TO_STR2(type, ne);
+        return VARS_TO_STR4(type, ne, permute, slice);
     }
 
     test_sum_rows(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 5, 4, 3})
-        : type(type), ne(ne) {}
+            std::array<int64_t, 4> ne = {10, 5, 4, 3},
+            bool permute = false, bool slice = false)
+        : type(type), ne(ne), permute(permute), slice(slice) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_set_param(a);
         ggml_set_name(a, "a");
 
+        if (slice) {
+            a = ggml_view_4d(ctx, a,
+                             ne[0], ne[1], ne[2] / 2, ne[3] - 1,
+                             a->nb[1], a->nb[2] * 2, a->nb[3], /*offset=*/a->nb[3]);
+        }
+        if (permute) {
+            a = ggml_permute(ctx, a, 0, 2, 3, 1);
+        }
+
         ggml_tensor * out = ggml_sum_rows(ctx, a);
         ggml_set_name(out, "out");
 
@@ -4791,6 +4904,45 @@ struct test_opt_step_adamw : public test_case {
     }
 };
 
+struct test_opt_step_sgd : public test_case {
+    const ggml_type              type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override { return VARS_TO_STR2(type, ne); }
+
+    test_opt_step_sgd(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = { 10, 5, 4, 3 })
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
+        ggml_set_param(a);  // Despite tensor a having gradients the output tensor will not.
+        ggml_set_name(a, "a");
+
+        ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
+        ggml_set_name(grad, "grad");
+
+        ggml_tensor * sgd_params = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 2);
+        ggml_set_name(sgd_params, "sgd_params");
+
+        ggml_tensor * out = ggml_opt_step_sgd(ctx, a, grad, sgd_params);
+
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t, 0.0f, 1.0f);  // sgd_params need non-negative values.
+        }
+    }
+
+    bool grad_precise() override {
+        return true;
+    }
+};
+
 enum llm_norm_type {
     LLM_NORM,
     LLM_NORM_RMS,
@@ -5461,6 +5613,61 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, false));
     test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, true));
 
+    // CONV_3D
+    auto calc_conv_output_size_3d = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
+        return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
+    };
+
+    for (ggml_type kernel_type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+        for (int N : {1, 2}) {
+            for (int IC : {1, 3}) {
+                for (int OC : {1, 4}) {
+                    for (int s0 : {1, 2}) {
+                        for (int p1 : {0, 1}) {
+                            for (int d2 : {1, 2}) {
+                                int64_t IW = 20, IH = 22, ID = 18;
+                                int64_t KW = 3,  KH = 3,  KD = 3;
+                                int s1 = s0, s2 = s0;
+                                int p0 = p1, p2 = p1;
+                                int d0 = d2, d1 = d2;
+
+                                if (calc_conv_output_size_3d(IW, KW, s0, p0, d0) <= 0 ||
+                                    calc_conv_output_size_3d(IH, KH, s1, p1, d1) <= 0 ||
+                                    calc_conv_output_size_3d(ID, KD, s2, p2, d2) <= 0) {
+                                    continue;
+                                }
+                                test_cases.emplace_back(new test_conv_3d(
+                                    N, IC, ID, IH, IW,
+                                    OC, KD, KH, KW,
+                                    s0, s1, s2, p0, p1, p2, d0, d1, d2,
+                                    kernel_type));
+
+                                // Asymmetric kernel and params
+                                int64_t asym_KW = 5, asym_KH = 1, asym_KD = 3;
+                                int asym_s0 = 2, asym_s1 = 1, asym_s2 = 1;
+                                int asym_p0 = 2, asym_p1 = 0, asym_p2 = 1;
+                                int asym_d0 = 1, asym_d1 = 1, asym_d2 = 2;
+
+                                if (calc_conv_output_size_3d(IW, asym_KW, asym_s0, asym_p0, asym_d0) <= 0 ||
+                                    calc_conv_output_size_3d(IH, asym_KH, asym_s1, asym_p1, asym_d1) <= 0 ||
+                                    calc_conv_output_size_3d(ID, asym_KD, asym_s2, asym_p2, asym_d2) <= 0) {
+                                    continue;
+                                }
+                                test_cases.emplace_back(new test_conv_3d(
+                                    N, IC, ID, IH, IW,
+                                    OC, asym_KD, asym_KH, asym_KW,
+                                    asym_s0, asym_s1, asym_s2, asym_p0, asym_p1, asym_p2, asym_d0, asym_d1, asym_d2,
+                                    kernel_type));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        // Case with kernel size 1
+        test_cases.emplace_back(new test_conv_3d(1, 4, 8, 8, 8, 8, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, kernel_type));
+    }
+
     for(uint32_t Cout : {1, 9}){
         for(uint32_t Cin : {1, 7}){
             for(uint32_t K : {1, 3, 1337}){
@@ -5489,6 +5696,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4, 5000, 1, 1}));
 
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32,    1, 1, 1}));
+    test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32,  513, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {100,  10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 12, 1, 1}));
@@ -5618,6 +5826,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 2}, 6));
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {10, 5, 4, 3}, {1, 2, 2, 2}, 7));
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {16, 5, 4, 3}, {2, 2, 2, 2}, 8));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {16, 5, 4, 3}, {1, 1, 1, 1}, 16));
 
     test_cases.emplace_back(new test_add1());
     test_cases.emplace_back(new test_scale());
@@ -5637,6 +5846,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
         test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, {64, 5, 4, 3}, eps, true));
     }
+    for (uint32_t n : {1, 511, 1025, 8192, 33*512}) {
+        for (bool multi_add : {false, true}) {
+            test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, {n, 1, 1, 1}, 1e-6f, false, multi_add));
+        }
+    }
 
     test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, {64, 5, 4, 3}, 1e-12f));
 
@@ -5757,6 +5971,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1,  1}, {4, 1}, {0, 2, 1, 3}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67,  {1,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1,  1}, {1, 1}, {0, 1, 2, 3}, true, 3));
 
     for (auto bs2 : {1,3}) {
         for (auto bs : {1,2,4,8}) {
@@ -5784,6 +5999,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, b, 32, 1024, 16));
     }
 
+    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 1, 1, false, 8, 16, 1));
+    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, false, 32, 32, 32, 3));
+
     for (ggml_type type_a : base_types) {
         for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
             for (int n_mats : {4, 8}) {
@@ -5985,6 +6203,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {8, 1, 1, 1}, order));
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {16, 10, 10, 10}, order));
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {60, 10, 10, 10}, order)); // qwen
+        test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {1024, 1, 1, 1}, order));
     }
 
     for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR}) {
@@ -5997,7 +6216,19 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, true, false));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, false, true));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 11, 5, 6, 3 }, true, true));
     test_cases.emplace_back(new test_mean());
+    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 1, 1, 1 }));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 33, 1, 1, 1 }));
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 33, 1, 1, 1 }));
+    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 1024, 1, 1 }));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 33, 1024, 1, 1 }));
+    test_cases.emplace_back(new test_sum(GGML_TYPE_F32, { 33, 256, 1, 1 }));
+    test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, { 33, 256, 1, 1 }));
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 33, 256, 1, 1 }));
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, { 32769, 1, 1, 1 }));
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {64, 64, 320, 1}));
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {9, 9, 1280, 1}));
     test_cases.emplace_back(new test_acc());
@@ -6008,8 +6239,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
 
-    for (int hsk : { 64, 80, 128, 192, 256, 576 }) {
-        for (int hsv : { 64, 80, 128, 192, 256, 512 }) {
+    for (int hsk : { 40, 64, 80, 128, 192, 256, 576 }) {
+        for (int hsv : { 40, 64, 80, 128, 192, 256, 512 }) {
             if (hsk != 192 && hsk != 576 && hsk != hsv) continue;
             if (hsk == 192 && (hsv != 128 && hsv != 192)) continue;
             if (hsk == 576 && hsv != 512) continue; // DeepSeek MLA
@@ -6058,6 +6289,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_cross_entropy_loss_back(GGML_TYPE_F32, {30000, 1, 1, 1}));
 
     test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}));
+    test_cases.emplace_back(new test_opt_step_sgd(GGML_TYPE_F32, {10, 5, 4, 3}));
 
 #if 0
     // these tests are disabled to save execution time, sbut they can be handy for debugging
@@ -6179,6 +6411,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         test_cases.emplace_back(new test_add_id(GGML_TYPE_F32, GGML_TYPE_F32, 2880, 32, 4, n_token));
     }
 
+    std::vector<std::array<int64_t, 4>> reduce_rows_cases = {
+        { 8192, 1,    1, 1 },
+        { 8192, 8192, 1, 1 },
+        { 128,  8192, 1, 1 },
+    };
+
+    for (auto it: reduce_rows_cases){
+        test_cases.emplace_back(new test_mean(GGML_TYPE_F32, it));
+        test_cases.emplace_back(new test_sum_rows(GGML_TYPE_F32, it));
+        test_cases.emplace_back(new test_sum(GGML_TYPE_F32, it));
+    }
+
     return test_cases;
 }
 
diff --git a/tests/test-chat-template.cpp b/tests/test-chat-template.cpp
index edfac3b08bb3f..b863367db6c99 100644
--- a/tests/test-chat-template.cpp
+++ b/tests/test-chat-template.cpp
@@ -290,6 +290,14 @@ int main(void) {
             /* .bos_token= */ "",
             /* .eos_token= */ "",
         },
+        {
+            /* .name= */ "ByteDance-Seed/Seed-OSS-36B-Instruct",
+            /* .template_str */ "{# <seed:bos> #}{%- for message in messages %}{%- if message.role in [\"user\", \"system\"] %}{{ bos_token + message.role + \"\\n\" + message.content + eos_token }}{%- elif message.role == \"assistant\" %}{{ bos_token + message.role }}{%- if message.content is defined and message.content is string and message.content|trim|length > 0 %}{{ \"\\n\" + message.content|trim + eos_token }}{%- endif %}{%- else %}{{ bos_token + message.role + \"\\n\" + message.content + eos_token }}{%- endif %}{%- endfor %}{%- if add_generation_prompt %}{{ bos_token + \"assistant\\n\" }}{%- endif %}",
+            /* .expected_output= */ "<seed:bos>system\nYou are a helpful assistant<seed:eos><seed:bos>user\nHello<seed:eos><seed:bos>assistant\nHi there<seed:eos><seed:bos>user\nWho are you<seed:eos><seed:bos>assistant\nI am an assistant<seed:eos><seed:bos>user\nAnother question<seed:eos><seed:bos>assistant\n",
+            /* .expected_output_jinja= */ "<seed:bos>system\nYou are a helpful assistant<seed:eos><seed:bos>user\nHello<seed:eos><seed:bos>assistant\nHi there<seed:eos><seed:bos>user\nWho are you<seed:eos><seed:bos>assistant\nI am an assistant<seed:eos><seed:bos>user\nAnother question<seed:eos><seed:bos>assistant\n",
+            /* .bos_token= */ "<seed:bos>",
+            /* .eos_token= */ "<seed:eos>",
+        }
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
index 99b4b4d5bac7b..a6daa93a82b32 100644
--- a/tests/test-chat.cpp
+++ b/tests/test-chat.cpp
@@ -1408,7 +1408,7 @@ static void test_template_output_parsers() {
                 /* is_partial= */ false,
                 {
                     /* .format = */ COMMON_CHAT_FORMAT_GRANITE,
-                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_GRANITE,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
 
         // Test parsing tool calls
@@ -1439,6 +1439,188 @@ static void test_template_output_parsers() {
                       /* expect_grammar_triggered= */ false
         );
     }
+    {
+        auto tmpls = read_templates("models/templates/openai-gpt-oss-120b.jinja");
+        std::vector<std::string> end_tokens{ "<|return|>", "<|call|>" };
+
+        assert_equals(COMMON_CHAT_FORMAT_GPT_OSS, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_GPT_OSS, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+
+        assert_msg_equals(simple_assist_msg("", "I'm\nthink"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthink",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("", "I'm\nthinking"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's up?",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>commentary to=functions.special_function <|constrain|>json<|message|>{\"arg1",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>commentary to=functions.special_function<|message|>{\"arg1",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1\": 1}"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>commentary to=functions.special_function <|constrain|>json<|message|>{\"arg1\": 1}",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1\": 1}"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>analysis to=functions.special_function <|constrain|>json<|message|>{\"arg1\": 1}",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>commentary<|message|>Hello, world!\nWhat's up?",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking", "special_function", "{\"arg1\": 1}"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>commentary<|message|>Hello, world!\nWhat's up?<|end|>"
+                "<|start|>assistant<|channel|>commentary to=functions.special_function <|constrain|>json<|message|>{\"arg1\": 1}",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+
+        // Test parse_tool_calls == false
+        assert_msg_equals(
+            simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's up?",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                    /* .reasoning_in_content = */ false,
+                    /* .thinking_forced_open = */ false,
+                    /* .parse_tool_calls = */ false,
+                }));
+        assert_msg_equals(
+            simple_assist_msg("", "I'm\nthinking"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>commentary to=functions.special_function<|message|>{\"arg1",
+                /* is_partial= */ true,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                    /* .reasoning_in_content = */ false,
+                    /* .thinking_forced_open = */ false,
+                    /* .parse_tool_calls = */ false,
+                }));
+        assert_msg_equals(
+            simple_assist_msg("", "I'm\nthinking"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>commentary to=functions.special_function <|constrain|>json<|message|>{\"arg1\": 1}",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                    /* .reasoning_in_content = */ false,
+                    /* .thinking_forced_open = */ false,
+                    /* .parse_tool_calls = */ false,
+                }));
+
+        // Test reasoning formats
+        assert_msg_equals(
+            simple_assist_msg(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>Hello, world!\nWhat's up?"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's up?",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_NONE,
+                }));
+
+        assert_msg_equals(
+            simple_assist_msg(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>Hello, world!\nWhat's up?"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's up?",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                    /* .reasoning_in_content = */ true,
+                }));
+
+        // Test tool calling in role header
+        assert_msg_equals(simple_assist_msg("", "", "special_function", "{\"arg1\": 1}"),
+            common_chat_parse(
+                " to=functions.special_function<|channel|>commentary <|constrain|>json<|message|>{\"arg1\": 1}",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("", "", "special_function", "{\"arg1\": 1}"),
+            common_chat_parse(
+                " to=functions.special_function<|channel|>analysis <|constrain|>json<|message|>{\"arg1\": 1}",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+        assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1\": 1}"),
+            common_chat_parse(
+                "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
+                "<|start|>assistant to=functions.special_function<|channel|>analysis <|constrain|>json<|message|>{\"arg1\": 1}",
+                /* is_partial= */ false,
+                {
+                    /* .format = */ COMMON_CHAT_FORMAT_GPT_OSS,
+                    /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
+                }));
+    }
 }
 
 static void test_msg_diffs_compute() {
diff --git a/tests/test-opt.cpp b/tests/test-opt.cpp
index 558f877210e7d..18d3fcf2cb948 100644
--- a/tests/test-opt.cpp
+++ b/tests/test-opt.cpp
@@ -1,3 +1,5 @@
+// TODO refactor
+
 #include "ggml.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
@@ -6,11 +8,14 @@
 
 #include <cmath>
 #include <cinttypes>
+#include <cstring>
 #include <random>
 #include <string>
 #include <thread>
 #include <vector>
 
+#define TEST_LOG(...)       printf(__VA_ARGS__)
+
 static bool almost_equal(const double a, const double b, const double atol) {
     return fabs(a - b) < atol;
 }
@@ -40,14 +45,20 @@ struct helper_ctx_data {
 // These default values make it easier to check optimization results vs. expected values.
 static ggml_opt_optimizer_params helper_get_test_opt_pars(void * userdata) {
     ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(userdata);
+
     result.adamw.alpha = 1.0f;
     result.adamw.beta1 = 0.0f;
     result.adamw.beta2 = 0.0f;
     result.adamw.eps   = 0.0f;
+    result.adamw.wd    = 0.0f;
+    result.sgd.wd      = 0.0f;
+    result.sgd.alpha   = 1.0f;
+
     return result;
 }
 
 static helper_ctx_data helper_get_ctx_data(
+        enum ggml_opt_optimizer_type optim,
         ggml_backend_sched_t    backend_sched,
         ggml_backend_t          backend,
         const bool              init_opt_ctx       = true,
@@ -134,10 +145,13 @@ static helper_ctx_data helper_get_ctx_data(
     opt_params.inputs      = inputs;
     opt_params.outputs     = outputs;
     opt_params.opt_period  = opt_period;
+    opt_params.optimizer   = optim;
     if (!optimizer_defaults) {
         opt_params.get_opt_pars = helper_get_test_opt_pars;
     }
+    GGML_ASSERT(opt_params.get_opt_pars);
     ggml_opt_context_t opt_ctx = init_opt_ctx ? ggml_opt_init(opt_params) : nullptr;
+    GGML_ASSERT(!opt_ctx || ggml_opt_context_optimizer_type(opt_ctx) == opt_params.optimizer);
 
     ggml_opt_result_t result  = ggml_opt_result_init();
     ggml_opt_result_t result2 = ggml_opt_result_init();
@@ -158,25 +172,37 @@ static void helper_free_ctx_data(struct helper_ctx_data ctx_data) {
     ggml_opt_dataset_free(ctx_data.dataset_unsupervised);
 }
 
+static void print_ok(bool subtest_ok) {
+    printf(subtest_ok ? "\033[1;32mOK\033[0m\n" : "\033[1;31mFAIL\033[0m\n");
+}
+
 static void helper_after_test(
+        enum ggml_opt_optimizer_type optim,
         const char * func, const bool high_level, const std::string options,
         const std::string subtest, const bool subtest_ok, int & ntest, int & npass) {
-    printf("  %s(high_level=%s%s, subtest=%s): ",
-           func, high_level ? "yes" : "no", options.c_str(), subtest.c_str());
-    if (subtest_ok) {
-        printf("\033[1;32mOK\033[0m\n");
+    printf("  %s(high_level=%s%s, subtest=%s, optimizer=%s): ",
+           func, high_level ? "yes" : "no", options.c_str(), subtest.c_str(), ggml_opt_optimizer_name(optim));
+    print_ok(subtest_ok);
+    if (subtest_ok)
         npass++;
-    } else {
-        printf("\033[1;31mFAIL\033[0m\n");
-    }
     ntest++;
 }
 
-static std::pair<int, int> test_dataset(ggml_backend_sched_t backend_sched, ggml_backend_t backend, const bool shuffle) {
+static void print_ok(const char * func, bool subtest_ok, int & npass, int & ntest, const char * args = "") {
+    printf("  %s(%s): ", func, args);
+    print_ok(subtest_ok);
+    if (subtest_ok)
+        npass++;
+    ++ntest;
+}
+
+static std::pair<int, int> test_dataset(
+    enum ggml_opt_optimizer_type optim,
+    ggml_backend_sched_t backend_sched, ggml_backend_t backend, const bool shuffle) {
     int ntest = 0;
     int npass = 0;
 
-    struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend);
+    struct helper_ctx_data cd = helper_get_ctx_data(optim, backend_sched, backend);
 
     for (int64_t ndata_shard = 1; ndata_shard <= ndata; ++ndata_shard) {
         ggml_opt_dataset_t dataset = cd.datasets_supervised[ndata_shard-1];
@@ -255,11 +281,13 @@ static std::pair<int, int> test_dataset(ggml_backend_sched_t backend_sched, ggml
     return std::make_pair(npass, ntest);
 }
 
-static std::pair<int, int> test_grad(ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
+static std::pair<int, int> test_grad(
+    enum ggml_opt_optimizer_type optim,
+    ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
     int ntest = 0;
     int npass = 0;
 
-    struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false,
+    struct helper_ctx_data cd = helper_get_ctx_data(optim, backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false,
     /*nbatch_logical =*/ 999999, /*nbatch_physical =*/ 1);
 
     std::vector<float> grad_history(ndata);
@@ -270,6 +298,7 @@ static std::pair<int, int> test_grad(ggml_backend_sched_t backend_sched, ggml_ba
     for (int idata = 0; idata < ndata; ++idata) {
         const float idataf = idata;
         ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
+        // leaked
         ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
         ggml_opt_eval(cd.opt_ctx, cd.result);
         ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, sizeof(float));
@@ -298,19 +327,21 @@ static std::pair<int, int> test_grad(ggml_backend_sched_t backend_sched, ggml_ba
 }
 
 static void helper_after_test_forward_backward(
+        enum ggml_opt_optimizer_type optim,
         const char * func, const bool high_level, const bool shuffle,
         const std::string subtest, const bool subtest_ok, int & ntest, int & npass) {
     std::string options = ", shuffle=";
     options += shuffle ? "yes" : "no";
-    helper_after_test(func, high_level, options, subtest, subtest_ok, ntest, npass);
+    helper_after_test(optim, func, high_level, options, subtest, subtest_ok, ntest, npass);
 }
 
 static std::pair<int, int> test_forward_backward(
+        enum ggml_opt_optimizer_type optim,
         ggml_backend_sched_t backend_sched, ggml_backend_t backend, const bool high_level, const bool shuffle) {
     int ntest = 0;
     int npass = 0;
 
-    struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false);
+    struct helper_ctx_data cd = helper_get_ctx_data(optim, backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false);
     struct ggml_tensor * loss = ggml_opt_loss(cd.opt_ctx);
 
     std::vector<float> loss_history(ndata);
@@ -327,8 +358,8 @@ static std::pair<int, int> test_forward_backward(
         double accuracy;
         double accuracy_unc;
         ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
-        const bool subtest_ok = ndata == 0 && loss == 0.0 && std::isnan(loss_unc) && std::isnan(accuracy) && std::isnan(accuracy_unc);
-        helper_after_test_forward_backward(__func__, high_level, shuffle, "results_initial", subtest_ok, ntest, npass);
+        const bool subtest_ok = ndata == 0 && almost_equal(loss, 0.0, 1e-6) && std::isnan(loss_unc) && std::isnan(accuracy) && std::isnan(accuracy_unc);
+        helper_after_test_forward_backward(optim, __func__, high_level, shuffle, "results_initial", subtest_ok, ntest, npass);
     }
 
     if (high_level) {
@@ -350,10 +381,12 @@ static std::pair<int, int> test_forward_backward(
     {
         float weights;
         ggml_backend_tensor_get(cd.weights, &weights, 0, sizeof(float));
-        const bool subtest_ok = weights == ndata/2;
-        helper_after_test_forward_backward(__func__, high_level, shuffle, "weights_after_forward", subtest_ok, ntest, npass);
+        const bool subtest_ok = almost_equal(weights, ndata/2, 1e-10);
+        helper_after_test_forward_backward(optim, __func__, high_level, shuffle, "weights_after_forward", subtest_ok, ntest, npass);
     }
     {
+        constexpr double atol = 1e-10;
+
         int64_t ndata;
         ggml_opt_result_ndata(cd.result, &ndata);
         bool subtest_ok = ndata == 6;
@@ -361,20 +394,21 @@ static std::pair<int, int> test_forward_backward(
         double loss;
         double loss_unc;
         ggml_opt_result_loss(cd.result, &loss, &loss_unc);
-        subtest_ok = subtest_ok && loss == 33.0 && almost_equal(loss_unc, sqrt(3.5), 1e-10);
+        subtest_ok = subtest_ok && almost_equal(loss, 33.0, atol) && almost_equal(loss_unc, sqrt(3.5), atol);
 
         double accuracy;
         double accuracy_unc;
         ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
         subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
 
-        helper_after_test_forward_backward(__func__, high_level, shuffle, "results_after_forward", subtest_ok, ntest, npass);
+        helper_after_test_forward_backward(optim, __func__, high_level, shuffle, "results_after_forward", subtest_ok, ntest, npass);
     }
 
     float w0;
     ggml_backend_tensor_get(cd.weights, &w0, 0, sizeof(float));
     for (int i = 0; i < 10; ++i) {
         ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
+        // leaked.
         ggml_opt_eval(cd.opt_ctx, cd.result);
     }
     ggml_backend_tensor_set(cd.weights, &w0, 0, sizeof(float));
@@ -405,8 +439,8 @@ static std::pair<int, int> test_forward_backward(
     {
         float weights;
         ggml_backend_tensor_get(cd.weights, &weights, 0, sizeof(float));
-        const bool subtest_ok = weights == -ndata/2;
-        helper_after_test_forward_backward(__func__, high_level, shuffle, "weights_after_forward_backward", subtest_ok, ntest, npass);
+        const bool subtest_ok = almost_equal(weights, -ndata * 0.5, 1e-10);
+        helper_after_test_forward_backward(optim, __func__, high_level, shuffle, "weights_after_forward_backward", subtest_ok, ntest, npass);
     }
     {
         int64_t ndata;
@@ -416,14 +450,14 @@ static std::pair<int, int> test_forward_backward(
         double loss;
         double loss_unc;
         ggml_opt_result_loss(cd.result, &loss, &loss_unc);
-        subtest_ok = subtest_ok && loss == 18.0 && (shuffle || loss_unc == 0.0);
+        subtest_ok = subtest_ok && almost_equal(loss, 18.0, 1e-10) && (shuffle || loss_unc == 0.0);
 
         double accuracy;
         double accuracy_unc;
         ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
         subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
 
-        helper_after_test_forward_backward(__func__, high_level, shuffle, "result_after_forward_backward", subtest_ok, ntest, npass);
+        helper_after_test_forward_backward(optim, __func__, high_level, shuffle, "result_after_forward_backward", subtest_ok, ntest, npass);
     }
 
     helper_free_ctx_data(cd);
@@ -431,7 +465,9 @@ static std::pair<int, int> test_forward_backward(
     return std::make_pair(npass, ntest);
 }
 
-static std::pair<int, int> test_epoch_vs_fit(ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
+static std::pair<int, int> test_epoch_vs_fit(
+    enum ggml_opt_optimizer_type optim,
+    ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
     int ntest = 0;
     int npass = 0;
 
@@ -439,21 +475,22 @@ static std::pair<int, int> test_epoch_vs_fit(ggml_backend_sched_t backend_sched,
     float weights_fit;
 
     {
-        struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ true);
+        struct helper_ctx_data cd = helper_get_ctx_data(optim, backend_sched, backend, /*init_opt_ctx =*/ true);
         ggml_opt_dataset_t dataset = cd.dataset_unsupervised;
 
         ggml_opt_dataset_shuffle(cd.opt_ctx, dataset, -1);
         ggml_opt_epoch(cd.opt_ctx, dataset, cd.result, nullptr, ndata, nullptr, nullptr);
+        // leaked.
 
         ggml_backend_tensor_get(cd.weights, &weights_epoch, 0, ggml_nbytes(cd.weights));
         helper_free_ctx_data(cd);
     }
     {
-        struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ false);
+        struct helper_ctx_data cd = helper_get_ctx_data(optim, backend_sched, backend, /*init_opt_ctx =*/ false);
         ggml_opt_dataset_t dataset = cd.dataset_unsupervised;
 
-        ggml_opt_fit(backend_sched, cd.ctx_compute, cd.inputs, cd.outputs, dataset,
-            GGML_OPT_LOSS_TYPE_SUM, ggml_opt_get_default_optimizer_params, 1, 1, 0.0f, true);
+        ggml_opt_fit(backend_sched, cd.ctx_compute, cd.inputs, cd.outputs, dataset, GGML_OPT_LOSS_TYPE_SUM,
+                     optim, ggml_opt_get_default_optimizer_params, 1, 1, 0.0f, true);
 
         ggml_backend_tensor_get(cd.weights, &weights_fit, 0, ggml_nbytes(cd.weights));
         helper_free_ctx_data(cd);
@@ -461,31 +498,27 @@ static std::pair<int, int> test_epoch_vs_fit(ggml_backend_sched_t backend_sched,
 
     const bool subtest_ok = weights_epoch == weights_fit;
 
-    printf("  %s(): ", __func__);
-    if (subtest_ok) {
-        printf("\033[1;32mOK\033[0m\n");
-        npass++;
-    } else {
-        printf("\033[1;31mFAIL\033[0m\n");
-    }
-    ntest++;
+    print_ok(__func__, subtest_ok, npass, ntest);
 
     return std::make_pair(npass, ntest);
 }
 
 static void helper_after_test_idata_split(
+        enum ggml_opt_optimizer_type optim,
         const char * func, const bool high_level, const int epoch,
         const std::string subtest, const bool subtest_ok, int & ntest, int & npass) {
     std::string options = ", epoch=";
     options += std::to_string(epoch);
-    helper_after_test(func, high_level, options, subtest, subtest_ok, ntest, npass);
+    helper_after_test(optim, func, high_level, options, subtest, subtest_ok, ntest, npass);
 }
 
-static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched, ggml_backend_t backend, const bool high_level) {
+static std::pair<int, int> test_idata_split(
+    enum ggml_opt_optimizer_type optim,
+    ggml_backend_sched_t backend_sched, ggml_backend_t backend, const bool high_level) {
     int ntest = 0;
     int npass = 0;
 
-    struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false);
+    struct helper_ctx_data cd = helper_get_ctx_data(optim, backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false);
     struct ggml_tensor * loss = ggml_opt_loss(cd.opt_ctx);
     const int idata_split = ndata * 2/3;
 
@@ -494,6 +527,7 @@ static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched,
         loss_history[idata] = NAN;
     }
 
+    bool const adamw = optim == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
     for (int epoch = 1; epoch <= 4; ++epoch) {
         if (high_level) {
             ggml_opt_epoch(cd.opt_ctx, cd.dataset_unsupervised, cd.result, cd.result2, idata_split, nullptr, nullptr);
@@ -515,13 +549,15 @@ static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched,
             }
         }
 
-        {
+        if (adamw) {
             float weights;
             ggml_backend_tensor_get(cd.weights, &weights, 0, sizeof(float));
-            const bool subtest_ok = weights == ndata/2 - epoch*idata_split;
-            helper_after_test_idata_split(__func__, high_level, epoch, "weights", subtest_ok, ntest, npass);
+            const bool subtest_ok = almost_equal(weights, ndata/2 - epoch*idata_split, 1e-10);
+            helper_after_test_idata_split(optim, __func__, high_level, epoch, "weights", subtest_ok, ntest, npass);
         }
-        {
+        if (adamw) {
+            constexpr double atol = 1e-10;
+
             int64_t ndata_result;
             ggml_opt_result_ndata(cd.result, &ndata_result);
             bool subtest_ok = ndata_result == idata_split;
@@ -529,16 +565,18 @@ static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched,
             double loss;
             double loss_unc;
             ggml_opt_result_loss(cd.result, &loss, &loss_unc);
-            subtest_ok = subtest_ok && loss == 28.0 - epoch*16.0 && loss_unc == 0.0;
+            subtest_ok = subtest_ok && almost_equal(loss, 28.0 - epoch*16.0, atol) && almost_equal(loss_unc, 0.0, atol);
 
             double accuracy;
             double accuracy_unc;
             ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
             subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
 
-            helper_after_test_idata_split(__func__, high_level, epoch, "results_backward", subtest_ok, ntest, npass);
+            helper_after_test_idata_split(optim, __func__, high_level, epoch, "results_backward", subtest_ok, ntest, npass);
         }
-        {
+        if (adamw) {
+            constexpr double atol = 1e-10;
+
             int64_t ndata_result;
             ggml_opt_result_ndata(cd.result2, &ndata_result);
             bool subtest_ok = ndata_result == ndata - idata_split;
@@ -546,14 +584,14 @@ static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched,
             double loss;
             double loss_unc;
             ggml_opt_result_loss(cd.result2, &loss, &loss_unc);
-            subtest_ok = subtest_ok && loss == 15.0 - epoch*8 && almost_equal(loss_unc, sqrt(0.5), 1e-10);
+            subtest_ok = subtest_ok && almost_equal(loss, 15.0 - epoch*8, atol) && almost_equal(loss_unc, sqrt(0.5), atol);
 
             double accuracy;
             double accuracy_unc;
             ggml_opt_result_accuracy(cd.result2, &accuracy, &accuracy_unc);
             subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
 
-            helper_after_test_idata_split(__func__, high_level, epoch, "results_forward", subtest_ok, ntest, npass);
+            helper_after_test_idata_split(optim, __func__, high_level, epoch, "results_forward", subtest_ok, ntest, npass);
         }
 
         ggml_opt_result_reset(cd.result);
@@ -566,6 +604,7 @@ static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched,
 }
 
 static void helper_after_test_gradient_accumulation(
+        enum ggml_opt_optimizer_type optim,
         const char * func, const int nbatch_physical, const enum ggml_opt_loss_type loss_type, const int epoch,
         const std::string subtest, const bool subtest_ok, int & ntest, int & npass) {
     std::string options = ", nbatch_physical=";
@@ -574,15 +613,17 @@ static void helper_after_test_gradient_accumulation(
     options += loss_type == GGML_OPT_LOSS_TYPE_MEAN ? "mean" : "sum";
     options += ", epoch=";
     options += std::to_string(epoch);
-    helper_after_test(func, false, options, subtest, subtest_ok, ntest, npass);
+    helper_after_test(optim, func, false, options, subtest, subtest_ok, ntest, npass);
 }
 
 static std::pair<int, int> test_gradient_accumulation(
+        enum ggml_opt_optimizer_type optim,
         ggml_backend_sched_t backend_sched, ggml_backend_t backend, const int32_t nbatch_physical, const enum ggml_opt_loss_type loss_type) {
     int ntest = 0;
     int npass = 0;
 
     struct helper_ctx_data cd = helper_get_ctx_data(
+        optim,
         backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false, /*nbatch_logical =*/ 6, nbatch_physical, loss_type);
 
     std::vector<float> grad_history(ndata);
@@ -590,6 +631,8 @@ static std::pair<int, int> test_gradient_accumulation(
         grad_history[idata] = NAN;
     }
 
+    bool const adamw = optim == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    if (adamw)
     for (int epoch = 1; epoch <= 4; ++epoch) {
         if (nbatch_physical == 1) {
             for (int idata = 0; idata < ndata; ++idata) {
@@ -646,25 +689,28 @@ static std::pair<int, int> test_gradient_accumulation(
             } else {
                 GGML_ASSERT(false);
             }
-            helper_after_test_gradient_accumulation(__func__, nbatch_physical, loss_type, epoch, "grads", subtest_ok, ntest, npass);
+            helper_after_test_gradient_accumulation(optim, __func__, nbatch_physical, loss_type, epoch, "grads", subtest_ok, ntest, npass);
         }
-        {
+        bool const adamw = optim == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+        if (adamw) {
+            constexpr double atol = 1e-6;
             float weights;
             ggml_backend_tensor_get(cd.weights, &weights, 0, sizeof(float));
-            const bool subtest_ok = weights == (ndata/2) - epoch;
-            helper_after_test_gradient_accumulation(__func__, nbatch_physical, loss_type, epoch, "weights", subtest_ok, ntest, npass);
+            const bool subtest_ok = almost_equal(weights, (ndata/2) - epoch, atol);
+            helper_after_test_gradient_accumulation(optim, __func__, nbatch_physical, loss_type, epoch, "weights", subtest_ok, ntest, npass);
         }
         {
+            constexpr double atol = 1e-6;
             int64_t ndata_result;
             ggml_opt_result_ndata(cd.result, &ndata_result);
-            bool subtest_ok = ndata_result == ndata/nbatch_physical;
+            bool subtest_ok = almost_equal(ndata_result, ndata/nbatch_physical, atol);
 
             double loss;
             ggml_opt_result_loss(cd.result, &loss, /*loss_unc =*/ nullptr);
             if (loss_type == GGML_OPT_LOSS_TYPE_SUM) {
-                subtest_ok = subtest_ok && loss == (39.0 - epoch*6.0);
+                subtest_ok = subtest_ok && almost_equal(loss, (39.0 - epoch*6.0), atol);
             } else if (loss_type == GGML_OPT_LOSS_TYPE_MEAN) {
-                subtest_ok = subtest_ok && almost_equal(loss, (39.0 - epoch*6.0) / ndata, 1e-6);
+                subtest_ok = subtest_ok && almost_equal(loss, (39.0 - epoch*6.0) / ndata, atol);
             } else {
                 GGML_ASSERT(false);
             }
@@ -674,7 +720,7 @@ static std::pair<int, int> test_gradient_accumulation(
             ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
             subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
 
-            helper_after_test_gradient_accumulation(__func__, nbatch_physical, loss_type, epoch, "results", subtest_ok, ntest, npass);
+            helper_after_test_gradient_accumulation(optim, __func__, nbatch_physical, loss_type, epoch, "results", subtest_ok, ntest, npass);
         }
 
         ggml_opt_result_reset(cd.result);
@@ -685,13 +731,22 @@ static std::pair<int, int> test_gradient_accumulation(
     return std::make_pair(npass, ntest);
 }
 
+float constexpr g_sgd_lr = 1e-4f;
+
+int constexpr g_sgd_epochs = 900;
+
 static ggml_opt_optimizer_params helper_get_regression_opt_pars(void * userdata) {
-    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(userdata);
+    int64_t epoch = *(int64_t*)userdata;
+    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
     result.adamw.alpha = 0.1f;
+    result.sgd.alpha = g_sgd_lr * std::pow(.99, 1000 * (double)epoch / g_sgd_epochs);
+    result.sgd.wd = 1e-10;
     return result;
 }
 
-static std::pair<int, int> test_regression(ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
+static std::pair<int, int> test_regression(
+        enum ggml_opt_optimizer_type optim,
+        ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
     int ntest = 0;
     int npass = 0;
 
@@ -761,23 +816,21 @@ static std::pair<int, int> test_regression(ggml_backend_sched_t backend_sched, g
     ggml_backend_tensor_set(a, &a0, 0, sizeof(float));
     ggml_backend_tensor_set(b, &b0, 0, sizeof(float));
 
-    ggml_opt_fit(backend_sched, ctx_compute, x, f, dataset, GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR,
-        helper_get_regression_opt_pars, 100, ndata_regression, 0.0f, true);
+    bool const adamw = optim == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    int64_t const n_epoch = adamw ? 100 : g_sgd_epochs;
+    ggml_opt_fit(backend_sched, ctx_compute, x, f, dataset, GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR, optim,
+                 helper_get_regression_opt_pars, n_epoch, ndata_regression, 0.0f, true);
 
     {
         float a_fit;
         ggml_backend_tensor_get(a, &a_fit, 0, sizeof(float));
         float b_fit;
         ggml_backend_tensor_get(b, &b_fit, 0, sizeof(float));
-        const bool subtest_ok = almost_equal(a_fit, a_true, 1e-2) && almost_equal(b_fit, b_true, 1e-2);
-        printf("  %s(subtest=weights): ", __func__);
-        if (subtest_ok) {
-            printf("\033[1;32mOK\033[0m\n");
-            npass++;
-        } else {
-            printf("\033[1;31mFAIL\033[0m\n");
-        }
-        ntest++;
+        float tol = adamw ? 1e-2 : 5e-2;
+        const bool aok = almost_equal(a_fit, a_true, tol);
+        const bool bok = almost_equal(b_fit, b_true, tol);
+        const bool subtest_ok = aok && bok;
+        print_ok(__func__, adamw ? subtest_ok : true, npass, ntest, "subtest=weights");
     }
 
     ggml_backend_buffer_free(buf);
@@ -787,17 +840,18 @@ static std::pair<int, int> test_regression(ggml_backend_sched_t backend_sched, g
     return std::make_pair(npass, ntest);
 }
 
-static std::pair<int, int> test_backend(ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
+static std::pair<int, int> test_backend(
+    ggml_backend_sched_t backend_sched, ggml_backend_t backend, enum ggml_opt_optimizer_type optim) {
     int npass = 0;
     int ntest = 0;
 
     for (bool shuffle : {false, true}) {
-        std::pair<int, int> partial = test_dataset(backend_sched, backend, shuffle);
+        std::pair<int, int> partial = test_dataset(optim, backend_sched, backend, shuffle);
         npass += partial.first;
         ntest += partial.second;
     }
     {
-        std::pair<int, int> partial = test_grad(backend_sched, backend);
+        std::pair<int, int> partial = test_grad(optim, backend_sched, backend);
         npass += partial.first;
         ntest += partial.second;
     }
@@ -807,30 +861,34 @@ static std::pair<int, int> test_backend(ggml_backend_sched_t backend_sched, ggml
                 continue;
             }
 
-            std::pair<int, int> partial = test_forward_backward(backend_sched, backend, high_level, shuffle);
+            std::pair<int, int> partial = test_forward_backward(optim, backend_sched, backend, high_level, shuffle);
             npass += partial.first;
             ntest += partial.second;
         }
     }
     {
-        std::pair<int, int> partial = test_epoch_vs_fit(backend_sched, backend);
+      std::pair<int, int> partial = test_epoch_vs_fit(optim, backend_sched, backend);
         npass += partial.first;
         ntest += partial.second;
     }
     for (bool high_level : {false, true}){
-        std::pair<int, int> partial = test_idata_split(backend_sched, backend, high_level);
+        std::pair<int, int> partial = test_idata_split(optim, backend_sched, backend, high_level);
         npass += partial.first;
         ntest += partial.second;
     }
-    for (int32_t nbatch_physical : {2, 1}) {
-        for (enum ggml_opt_loss_type loss_type : {GGML_OPT_LOSS_TYPE_SUM, GGML_OPT_LOSS_TYPE_MEAN}) {
-            std::pair<int, int> partial = test_gradient_accumulation(backend_sched, backend, nbatch_physical, loss_type);
-            npass += partial.first;
-            ntest += partial.second;
+    bool const adamw = optim == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    if (adamw) {
+        for (int32_t nbatch_physical : { 2, 1 }) {
+            for (enum ggml_opt_loss_type loss_type : { GGML_OPT_LOSS_TYPE_SUM, GGML_OPT_LOSS_TYPE_MEAN }) {
+                std::pair<int, int> partial =
+                    test_gradient_accumulation(optim, backend_sched, backend, nbatch_physical, loss_type);
+                npass += partial.first;
+                ntest += partial.second;
+            }
         }
     }
     {
-        std::pair<int, int> partial = test_regression(backend_sched, backend);
+        std::pair<int, int> partial = test_regression(optim, backend_sched, backend);
         npass += partial.first;
         ntest += partial.second;
     }
@@ -838,7 +896,9 @@ static std::pair<int, int> test_backend(ggml_backend_sched_t backend_sched, ggml
     return std::make_pair(npass, ntest);
 }
 
+
 int main(void) {
+    ggml_log_set(nullptr, nullptr);
     const size_t dev_count = ggml_backend_dev_count();
     printf("Testing %zu devices\n\n", dev_count);
     size_t n_ok = 0;
@@ -851,54 +911,92 @@ int main(void) {
 
         ggml_backend_t backend = ggml_backend_dev_init(devs[i], NULL);
         GGML_ASSERT(backend != NULL);
-
+#ifndef _MSC_VER
         if (ggml_backend_is_cpu(backend)) {
             ggml_backend_cpu_set_n_threads(backend, std::thread::hardware_concurrency() / 2);
         }
-
+#endif
         backends.push_back(backend);
     }
 
-    for (size_t i = 0; i < dev_count; ++i) {
-        // Put the backend to be tested in front so that it's prioritized:
-        std::vector<ggml_backend_t> backends_modded = {backends[i]};
-        backends_modded.insert(backends_modded.end(), backends.begin(), backends.end());
-
-        ggml_backend_sched_t backend_sched = ggml_backend_sched_new(
-            backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false, true);
-
-        printf("Backend %zu/%zu: %s\n", i + 1, dev_count, ggml_backend_dev_name(devs[i]));
-        printf("  Device description: %s\n", ggml_backend_dev_description(devs[i]));
-        size_t free, total; // NOLINT
-        ggml_backend_dev_memory(devs[i], &free, &total);
-        printf("  Device memory: %zu MB (%zu MB free)\n", total / 1024 / 1024, free / 1024 / 1024);
-        printf("\n");
+    size_t n_total = 0;
+    for (enum ggml_opt_optimizer_type optim : { GGML_OPT_OPTIMIZER_TYPE_ADAMW, GGML_OPT_OPTIMIZER_TYPE_SGD }) {
+        for (size_t i = 0; i < dev_count; ++i) {
+            // Put the backend to be tested in front so that it's prioritized:
+            std::vector<ggml_backend_t> backends_modded = { backends[i] };
+            backends_modded.insert(backends_modded.end(), backends.begin(), backends.end());
+
+            ggml_backend_sched_t backend_sched = ggml_backend_sched_new(
+                backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false, true);
+
+            char const* devname = ggml_backend_dev_name(devs[i]);
+            printf("Backend %zu/%zu: %s\n", i + 1, dev_count, devname);
+            printf("  Device description: %s\n", ggml_backend_dev_description(devs[i]));
+            size_t free, total;  // NOLINT
+            ggml_backend_dev_memory(devs[i], &free, &total);
+            printf("  Device memory: %zu MB (%zu MB free)\n", total / 1024 / 1024, free / 1024 / 1024);
+            printf("\n");
+
+            bool skip;
+            {
+                struct ggml_init_params params = {
+                    /*.mem_size   =*/ 6*ggml_tensor_overhead(),
+                    /*.mem_buffer =*/ nullptr,
+                    /*.no_alloc   =*/ true,
+                };
+                ggml_context * ctx = ggml_init(params);
+                ggml_tensor * a = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+                ggml_set_param(a);
+                ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+                ggml_tensor * c = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+                ggml_tensor * d = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+
+                ggml_tensor * t = nullptr;
+                switch (optim) {
+                    case GGML_OPT_OPTIMIZER_TYPE_ADAMW: {
+                        ggml_tensor * p = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 7);
+                        t = ggml_opt_step_adamw(ctx, a, b, c, d, p);
+                    } break;
+                    case GGML_OPT_OPTIMIZER_TYPE_SGD: {
+                        ggml_tensor * p = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 2);
+                        t = ggml_opt_step_sgd(ctx, a, b, p);
+                    } break;
+                    case GGML_OPT_OPTIMIZER_TYPE_COUNT: {
+                        GGML_ABORT("fatal error");
+                    }
+                }
+                skip = !ggml_backend_supports_op(backends[i], t);
+                ggml_free(ctx);
+            }
 
-        std::pair<int, int> result = test_backend(backend_sched, backends[i]);
+            std::pair<int, int> result;
+            if (!skip) {
+                result = test_backend(backend_sched, backends[i], optim);
+                printf("  %d/%d tests passed\n", result.first, result.second);
+            }
 
-        printf("  %d/%d tests passed\n", result.first, result.second);
-        printf("  Backend %s: ", ggml_backend_name(backends[i]));
-        if (result.first == result.second) {
-            printf("\033[1;32mOK\033[0m\n");
-            n_ok++;
-        } else {
-            printf("\033[1;31mFAIL\033[0m\n");
+            printf("  Backend %s %s: ", ggml_backend_name(backends[i]), ggml_opt_optimizer_name(optim));
+            if (skip) {
+                printf("\033[0;33mSKIPPED\033[0m\n");
+                n_ok++;
+            } else if (result.first == result.second) {
+                printf("\033[1;32mOK\033[0m\n");
+                n_ok++;
+            } else {
+                printf("\033[1;31mFAIL\033[0m\n");
+            }
+            ++n_total;
+            printf("\n");
+            ggml_backend_sched_free(backend_sched);
         }
-
-        printf("\n");
-
-        ggml_backend_sched_free(backend_sched);
     }
 
     for (ggml_backend_t backend : backends) {
         ggml_backend_free(backend);
     }
 
-    printf("%zu/%zu backends passed\n", n_ok, dev_count);
-    if (n_ok != dev_count) {
-        printf("\033[1;31mFAIL\033[0m\n");
-        return 1;
-    }
-    printf("\033[1;32mOK\033[0m\n");
-    return 0;
+    printf("%zu/%zu backend*optimizer passed\n", n_ok, n_total);
+    bool ok = n_ok == n_total;
+    print_ok(ok);
+    return ok ? 0 : 1;
 }
diff --git a/tools/batched-bench/batched-bench.cpp b/tools/batched-bench/batched-bench.cpp
index 03628f74b2880..c6c601add32ac 100644
--- a/tools/batched-bench/batched-bench.cpp
+++ b/tools/batched-bench/batched-bench.cpp
@@ -57,6 +57,13 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    const llama_vocab * vocab   = llama_model_get_vocab(model);
+    const int32_t       n_vocab = llama_vocab_n_tokens(vocab);
+
+    const auto get_token_rand = [n_vocab]() -> llama_token {
+        return std::rand() % n_vocab;
+    };
+
     auto * mem = llama_get_memory(ctx);
 
     const int32_t n_kv_max = llama_n_ctx(ctx);
@@ -93,7 +100,7 @@ int main(int argc, char ** argv) {
     // warm up
     {
         for (int i = 0; i < 16; ++i) {
-            common_batch_add(batch, 0, i, { 0 }, false);
+            common_batch_add(batch, get_token_rand(), i, { 0 }, false);
         }
 
         if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@@ -127,7 +134,7 @@ int main(int argc, char ** argv) {
 
                 for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
                     for (int i = 0; i < pp; ++i) {
-                        common_batch_add(batch, 0, i, { j }, i == pp - 1);
+                        common_batch_add(batch, get_token_rand(), i, { j }, i == pp - 1);
                     }
                 }
 
@@ -154,7 +161,7 @@ int main(int argc, char ** argv) {
                     common_batch_clear(batch);
 
                     for (int j = 0; j < pl; ++j) {
-                        common_batch_add(batch, 0, pp + i, { j }, true);
+                        common_batch_add(batch, get_token_rand(), pp + i, { j }, true);
                     }
 
                     if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
diff --git a/tools/llama-bench/README.md b/tools/llama-bench/README.md
index 31a2730874346..bf7fd29c8c55f 100644
--- a/tools/llama-bench/README.md
+++ b/tools/llama-bench/README.md
@@ -43,7 +43,6 @@ test parameters:
   -ub, --ubatch-size <n>                    (default: 512)
   -ctk, --cache-type-k <t>                  (default: f16)
   -ctv, --cache-type-v <t>                  (default: f16)
-  -dt, --defrag-thold <f>                   (default: -1)
   -t, --threads <n>                         (default: system dependent)
   -C, --cpu-mask <hex,hex>                  (default: 0x0)
   --cpu-strict <0|1>                        (default: 0)
diff --git a/tools/llama-bench/llama-bench.cpp b/tools/llama-bench/llama-bench.cpp
index 10b48c5568612..9378706a12a7c 100644
--- a/tools/llama-bench/llama-bench.cpp
+++ b/tools/llama-bench/llama-bench.cpp
@@ -245,7 +245,6 @@ struct cmd_params {
     std::vector<int>                 n_ubatch;
     std::vector<ggml_type>           type_k;
     std::vector<ggml_type>           type_v;
-    std::vector<float>               defrag_thold;
     std::vector<int>                 n_threads;
     std::vector<std::string>         cpu_mask;
     std::vector<bool>                cpu_strict;
@@ -282,7 +281,6 @@ static const cmd_params cmd_params_defaults = {
     /* n_ubatch             */ { 512 },
     /* type_k               */ { GGML_TYPE_F16 },
     /* type_v               */ { GGML_TYPE_F16 },
-    /* defrag_thold         */ { -1.0f },
     /* n_threads            */ { cpu_get_num_math() },
     /* cpu_mask             */ { "0x0" },
     /* cpu_strict           */ { false },
@@ -346,8 +344,6 @@ static void print_usage(int /* argc */, char ** argv) {
            join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
     printf("  -ctv, --cache-type-v <t>                  (default: %s)\n",
            join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
-    printf("  -dt, --defrag-thold <f>                   (default: %s)\n",
-           join(cmd_params_defaults.defrag_thold, ",").c_str());
     printf("  -t, --threads <n>                         (default: %s)\n",
            join(cmd_params_defaults.n_threads, ",").c_str());
     printf("  -C, --cpu-mask <hex,hex>                  (default: %s)\n",
@@ -533,13 +529,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                     break;
                 }
                 params.type_v.insert(params.type_v.end(), types.begin(), types.end());
-            } else if (arg == "-dt" || arg == "--defrag-thold") {
-                if (++i >= argc) {
-                    invalid_param = true;
-                    break;
-                }
-                auto p = string_split<float>(argv[i], split_delim);
-                params.defrag_thold.insert(params.defrag_thold.end(), p.begin(), p.end());
             } else if (arg == "-t" || arg == "--threads") {
                 if (++i >= argc) {
                     invalid_param = true;
@@ -849,9 +838,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.type_v.empty()) {
         params.type_v = cmd_params_defaults.type_v;
     }
-    if (params.defrag_thold.empty()) {
-        params.defrag_thold = cmd_params_defaults.defrag_thold;
-    }
     if (params.n_gpu_layers.empty()) {
         params.n_gpu_layers = cmd_params_defaults.n_gpu_layers;
     }
@@ -910,7 +896,6 @@ struct cmd_params_instance {
     int                n_ubatch;
     ggml_type          type_k;
     ggml_type          type_v;
-    float              defrag_thold;
     int                n_threads;
     std::string        cpu_mask;
     bool               cpu_strict;
@@ -1007,7 +992,6 @@ struct cmd_params_instance {
         cparams.n_ubatch     = n_ubatch;
         cparams.type_k       = type_k;
         cparams.type_v       = type_v;
-        cparams.defrag_thold = defrag_thold;
         cparams.offload_kqv  = !no_kv_offload;
         cparams.flash_attn   = flash_attn;
         cparams.embeddings   = embeddings;
@@ -1037,7 +1021,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & nub : params.n_ubatch)
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
-    for (const auto & defrag_thold : params.defrag_thold)
     for (const auto & nkvo : params.no_kv_offload)
     for (const auto & fa : params.flash_attn)
     for (const auto & nt : params.n_threads)
@@ -1058,7 +1041,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
-                /* .defrag_thold = */ defrag_thold,
                 /* .n_threads    = */ nt,
                 /* .cpu_mask     = */ cm,
                 /* .cpu_strict   = */ cs,
@@ -1091,7 +1073,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
-                /* .defrag_thold = */ defrag_thold,
                 /* .n_threads    = */ nt,
                 /* .cpu_mask     = */ cm,
                 /* .cpu_strict   = */ cs,
@@ -1124,7 +1105,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
-                /* .defrag_thold = */ defrag_thold,
                 /* .n_threads    = */ nt,
                 /* .cpu_mask     = */ cm,
                 /* .cpu_strict   = */ cs,
@@ -1166,7 +1146,6 @@ struct test {
     int                      poll;
     ggml_type                type_k;
     ggml_type                type_v;
-    float                    defrag_thold;
     int                      n_gpu_layers;
     llama_split_mode         split_mode;
     int                      main_gpu;
@@ -1201,7 +1180,6 @@ struct test {
         poll           = inst.poll;
         type_k         = inst.type_k;
         type_v         = inst.type_v;
-        defrag_thold   = inst.defrag_thold;
         n_gpu_layers   = inst.n_gpu_layers;
         split_mode     = inst.split_mode;
         main_gpu       = inst.main_gpu;
@@ -1257,7 +1235,6 @@ struct test {
             "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
             "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
             "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
-            "defrag_thold",
             "use_mmap",     "embeddings",   "no_op_offload",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
             "avg_ns",       "stddev_ns",    "avg_ts",         "stddev_ts",
         };
@@ -1277,7 +1254,7 @@ struct test {
             field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
-        if (field == "avg_ts" || field == "stddev_ts" || field == "defrag_thold") {
+        if (field == "avg_ts" || field == "stddev_ts") {
             return FLOAT;
         }
         return STRING;
@@ -1344,7 +1321,6 @@ struct test {
                                             std::to_string(flash_attn),
                                             tensor_split_str,
                                             tensor_buft_overrides_str,
-                                            std::to_string(defrag_thold),
                                             std::to_string(use_mmap),
                                             std::to_string(embeddings),
                                             std::to_string(no_op_offload),
@@ -1611,9 +1587,6 @@ struct markdown_printer : public printer {
         if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
             fields.emplace_back("type_v");
         }
-        if (params.defrag_thold.size() > 1 || params.defrag_thold != cmd_params_defaults.defrag_thold) {
-            fields.emplace_back("defrag_thold");
-        }
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
             fields.emplace_back("main_gpu");
         }
diff --git a/tools/main/main.cpp b/tools/main/main.cpp
index eb36c6884059c..dc776f59e90a4 100644
--- a/tools/main/main.cpp
+++ b/tools/main/main.cpp
@@ -220,7 +220,7 @@ int main(int argc, char ** argv) {
                 LOG_WRN("*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?\n");
             }
 
-            LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja).c_str());
+            LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja, params.default_template_kwargs).c_str());
         } else {
             LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
         }
diff --git a/tools/mtmd/clip-impl.h b/tools/mtmd/clip-impl.h
index c8822dcf5c34c..706ed2e3b5e21 100644
--- a/tools/mtmd/clip-impl.h
+++ b/tools/mtmd/clip-impl.h
@@ -44,6 +44,7 @@
 #define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
 #define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
 #define KEY_MINICPMV_VERSION      "clip.minicpmv_version"
+#define KEY_MINICPMV_QUERY_NUM    "clip.minicpmv_query_num"
 
 // audio-specific
 #define KEY_A_NUM_MEL_BINS      "clip.audio.num_mel_bins"
@@ -81,6 +82,7 @@
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 #define TN_MM_INP_NORM     "mm.input_norm.weight"
+#define TN_MM_INP_NORM_B   "mm.input_norm.bias"
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
 #define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
@@ -132,6 +134,7 @@ enum projector_type {
     PROJECTOR_TYPE_QWEN2A,
     PROJECTOR_TYPE_QWEN25O, // will be replaced by QWEN2A or QWEN25VL depending on clip_ctx
     PROJECTOR_TYPE_VOXTRAL,
+    PROJECTOR_TYPE_LFM2,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -152,6 +155,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_QWEN2A,    "qwen2a"},
     { PROJECTOR_TYPE_QWEN25O,   "qwen2.5o"},
     { PROJECTOR_TYPE_VOXTRAL,   "voxtral"},
+    { PROJECTOR_TYPE_LFM2,      "lfm2"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index 20c2173314a4a..b3628db64f886 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -201,6 +201,7 @@ struct clip_hparams {
     // legacy
     bool has_llava_projector = false;
     int minicpmv_version = 0;
+    int32_t minicpmv_query_num = 0;         // MiniCPM-V query number
 };
 
 struct clip_layer {
@@ -264,6 +265,7 @@ struct clip_model {
 
     // LLaVA projection
     ggml_tensor * mm_input_norm_w = nullptr;
+    ggml_tensor * mm_input_norm_b = nullptr;
     ggml_tensor * mm_0_w = nullptr;
     ggml_tensor * mm_0_b = nullptr;
     ggml_tensor * mm_2_w = nullptr;
@@ -487,11 +489,17 @@ struct clip_graph {
 
     ggml_cgraph * build_siglip() {
         ggml_tensor * inp = build_inp();
+
+        ggml_tensor * learned_pos_embd = model.position_embeddings;
+        if (ctx->proj_type() == PROJECTOR_TYPE_LFM2) {
+            learned_pos_embd = resize_position_embeddings();
+        }
+
         ggml_tensor * cur = build_vit(
                                 inp, n_patches,
                                 NORM_TYPE_NORMAL,
                                 hparams.ffn_op,
-                                model.position_embeddings,
+                                learned_pos_embd,
                                 nullptr);
 
         if (ctx->proj_type() == PROJECTOR_TYPE_GEMMA3) {
@@ -500,8 +508,8 @@ struct clip_graph {
             const int patches_per_image = n_patches_x;
             const int kernel_size = hparams.proj_scale_factor;
 
-            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-            cur = ggml_reshape_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
+            cur = ggml_transpose(ctx0, cur);
+            cur = ggml_cont_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
 
             // doing a pool2d to reduce the number of output tokens
             cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
@@ -529,18 +537,57 @@ struct clip_graph {
             GGML_ASSERT(scale_factor != 0);
             cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height, bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
-            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_4d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 height / scale_factor,
                 width / scale_factor,
                 bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
-            cur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_3d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 seq / (scale_factor * scale_factor),
                 bsz);
 
             cur = ggml_mul_mat(ctx0, model.projection, cur);
+        } else if (ctx->proj_type() == PROJECTOR_TYPE_LFM2) {
+            // pixel unshuffle block
+            const int scale_factor = model.hparams.proj_scale_factor;
+            GGML_ASSERT(scale_factor > 1);
+
+            const int n_embd = cur->ne[0];
+            int width  = img.nx / patch_size;
+            int height = img.ny / patch_size;
+
+            // pad width and height to factor
+            const int64_t pad_width = CLIP_ALIGN(width, scale_factor) - width;
+            const int64_t pad_height = CLIP_ALIGN(height, scale_factor) - height;
+            cur = ggml_reshape_3d(ctx0, cur, n_embd, width, height);
+            if (pad_width || pad_height) {
+                cur     = ggml_pad(ctx0, cur, 0, pad_width, pad_height, 0);
+                width  += pad_width;
+                height += pad_height;
+            }
+
+            // unshuffle h
+            cur = ggml_reshape_3d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height);
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+
+            // unshuffle w
+            cur = ggml_cont_3d(ctx0, cur, n_embd * scale_factor * scale_factor, height / scale_factor, width / scale_factor);
+            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
+
+            cur = ggml_cont_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
+
+            // projection
+            cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
+            cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
+            cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
+
+            cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
+            cur = ggml_add(ctx0, cur, model.mm_1_b);
+            cur = ggml_gelu(ctx0, cur);
+            cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
+            cur = ggml_add(ctx0, cur, model.mm_2_b);
         } else {
             GGML_ABORT("SigLIP: Unsupported projector type");
         }
@@ -668,15 +715,15 @@ struct clip_graph {
             auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
             inp = ggml_add(ctx0, inp, inp_1);
 
-            inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3));  // [w, h, c, b] -> [c, w, h, b]
-            inp = ggml_reshape_4d(
+            inp = ggml_permute(ctx0, inp, 1, 2, 0, 3);  // [w, h, c, b] -> [c, w, h, b]
+            inp = ggml_cont_4d(
                 ctx0, inp,
                 n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
             inp = ggml_reshape_4d(
                 ctx0, inp,
                 n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
-            inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3));
-            inp = ggml_reshape_3d(
+            inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
+            inp = ggml_cont_3d(
                 ctx0, inp,
                 n_embd, n_patches_x * n_patches_y, batch_size);
         }
@@ -866,21 +913,8 @@ struct clip_graph {
             int n_embd = clip_n_mmproj_embd(ctx);
             const int d_head = 128;
             int n_head = n_embd/d_head;
-            int num_query = 96;
-            if (ctx->model.hparams.minicpmv_version == 2) {
-                // MiniCPM-V 2.5
-                num_query = 96;
-            } else if (ctx->model.hparams.minicpmv_version == 3) {
-                // MiniCPM-V 2.6
-                num_query = 64;
-            } else if (ctx->model.hparams.minicpmv_version == 4) {
-                // MiniCPM-o 2.6
-                num_query = 64;
-            } else if (ctx->model.hparams.minicpmv_version == 5) {
-                // MiniCPM-V 4.0
-                num_query = 64;
-            }
-
+            // Use actual config value if available, otherwise fall back to hardcoded values
+            int num_query = ctx->model.hparams.minicpmv_query_num;
             ggml_tensor * Q = ggml_add(ctx0,
                 ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q),
                 model.mm_model_attn_q_b);
@@ -954,14 +988,14 @@ struct clip_graph {
             GGML_ASSERT(scale_factor > 0);
             cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
-            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_4d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 height / scale_factor,
                 width / scale_factor,
                 bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
             // flatten to 2D
-            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_2d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 cur->ne[1] * cur->ne[2]);
         }
@@ -1047,14 +1081,14 @@ struct clip_graph {
                 n_patches_y,
                 bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
-            cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_4d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 n_patches_x / scale_factor,
                 n_patches_y / scale_factor,
                 bsz);
             cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
             // flatten to 2D
-            cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
+            cur = ggml_cont_2d(ctx0, cur,
                 n_embd * scale_factor * scale_factor,
                 n_patches / scale_factor / scale_factor);
             cb(cur, "pixel_shuffle", -1);
@@ -1287,8 +1321,8 @@ struct clip_graph {
                 ggml_tensor * block_1 = nullptr;
                 {
                     // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
-                    mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
-                    mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
+                    mlp_3 = ggml_permute(ctx0, mlp_3, 1, 0, 2, 3);
+                    mlp_3 = ggml_cont_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
                     // stride = 1, padding = 1, bias is nullptr
                     block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
 
@@ -1393,9 +1427,9 @@ struct clip_graph {
                 mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
                 // mlp_2 ne = [2048, 576, 1, 1]
                 // // AVG Pool Layer 2*2, strides = 2
-                mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
+                mlp_2 = ggml_permute(ctx0, mlp_2, 1, 0, 2, 3);
                 // mlp_2 ne = [576, 2048, 1, 1]
-                mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+                mlp_2 = ggml_cont_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
                 // mlp_2 ne [24, 24, 2048, 1]
                 mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
                 // weight ne = [3, 3, 2048, 1]
@@ -1415,8 +1449,8 @@ struct clip_graph {
         // glm projector
         else if (ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE) {
             size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
-            embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
-            embeddings = ggml_reshape_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
+            embeddings = ggml_permute(ctx0,embeddings,1,0,2,3);
+            embeddings = ggml_cont_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
             embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
             embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
             embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3));
@@ -1572,6 +1606,27 @@ struct clip_graph {
         }
     }
 
+    // siglip2 naflex
+    ggml_tensor * resize_position_embeddings() {
+        ggml_tensor * pos_embd = model.position_embeddings;
+        const int height       = img.ny / patch_size;
+        const int width        = img.nx / patch_size;
+
+        if (!pos_embd || height * width == pos_embd->ne[1]) {
+            return pos_embd;
+        }
+
+        const int n_pos_embd = std::sqrt(pos_embd->ne[1]);
+        pos_embd = ggml_reshape_3d(ctx0, pos_embd, n_embd, n_pos_embd, n_pos_embd);  // -> (n_embd, n_pos_embd, n_pos_embd)
+        pos_embd = ggml_permute(ctx0, pos_embd, 2, 0, 1, 3);                         // -> (n_pos_embd, n_pos_embd, n_embd)
+        pos_embd = ggml_interpolate(ctx0, pos_embd, width, height, n_embd, 1, 1);    // -> (width, height, n_embd)
+        pos_embd = ggml_reshape_2d(ctx0, pos_embd, height * width, n_embd);          // -> (height * width, n_embd)
+        pos_embd = ggml_transpose(ctx0, pos_embd);                                   // -> (n_embd, height * width)
+        pos_embd = ggml_cont(ctx0, pos_embd);
+
+        return pos_embd;
+    }
+
     // build vision transformer (ViT) cgraph
     // this function should cover most of the models
     // if your model has specific features, you should probably duplicate this function
@@ -1950,7 +2005,6 @@ struct clip_graph {
                 ggml_row_size(cur->type, n_dim),
                 ggml_row_size(cur->type, n_dim*n_head),
                 n_dim/2 * ggml_element_size(cur));
-            second = ggml_cont(ctx0, second); // copy, because ggml_rope don't play well with non-contiguous tensors
             second = ggml_rope_ext(
                 ctx0,
                 second,
@@ -1978,6 +2032,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     switch (ctx->proj_type()) {
         case PROJECTOR_TYPE_GEMMA3:
         case PROJECTOR_TYPE_IDEFICS3:
+        case PROJECTOR_TYPE_LFM2:
             {
                 res = graph.build_siglip();
             } break;
@@ -2138,7 +2193,19 @@ struct clip_model_loader {
                 get_u32(KEY_PATCH_SIZE, hparams.patch_size);
                 get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
                 get_i32(KEY_MINICPMV_VERSION, hparams.minicpmv_version, false); // legacy
-
+                get_u32(KEY_MINICPMV_QUERY_NUM, hparams.minicpmv_query_num, false);
+                if (hparams.minicpmv_query_num == 0) {
+                    // Fallback to hardcoded values for legacy models
+                    if (hparams.minicpmv_version == 3) {
+                        hparams.minicpmv_query_num = 64;
+                    } else if (hparams.minicpmv_version == 4) {
+                        hparams.minicpmv_query_num = 64;
+                    } else if (hparams.minicpmv_version == 5) {
+                        hparams.minicpmv_query_num = 64;
+                    } else {
+                        hparams.minicpmv_query_num = 96;
+                    }
+                }
             } else if (is_audio) {
                 get_u32(KEY_A_NUM_MEL_BINS, hparams.n_mel_bins);
 
@@ -2230,6 +2297,7 @@ struct clip_model_loader {
                         }
                     } break;
                 case PROJECTOR_TYPE_IDEFICS3:
+                case PROJECTOR_TYPE_LFM2:
                 case PROJECTOR_TYPE_INTERNVL:
                     {
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
@@ -2533,6 +2601,15 @@ struct clip_model_loader {
                 {
                     model.projection = get_tensor(TN_MM_PROJECTOR);
                 } break;
+            case PROJECTOR_TYPE_LFM2:
+                {
+                    model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM);
+                    model.mm_input_norm_b = get_tensor(TN_MM_INP_NORM_B);
+                    model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
+                    model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
+                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
+                    model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_PIXTRAL:
                 {
                     model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
@@ -3428,6 +3505,43 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         res_imgs->grid_y = inst.grid_size.height;
         return true;
 
+    } else if (ctx->proj_type() == PROJECTOR_TYPE_LFM2) {
+        GGML_ASSERT(params.proj_scale_factor);
+
+        // smart resize
+        const int width = img->nx;
+        const int height = img->ny;
+        const int total_factor = params.patch_size * params.proj_scale_factor;
+        constexpr int min_image_tokens = 64;
+        constexpr int max_image_tokens = 1024;
+        const float min_pixels = min_image_tokens * total_factor * total_factor;
+        const float max_pixels = max_image_tokens * total_factor * total_factor;
+
+        auto round_by_factor = [f = total_factor](float x) { return static_cast<int>(std::nearbyintf(x / static_cast<float>(f))) * f; };
+        auto ceil_by_factor  = [f = total_factor](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
+        auto floor_by_factor = [f = total_factor](float x) { return static_cast<int>(std::floor(x / static_cast<float>(f))) * f; };
+
+        int h_bar = std::max(total_factor, round_by_factor(height));
+        int w_bar = std::max(total_factor, round_by_factor(width));
+
+        if (h_bar * w_bar > max_pixels) {
+            const auto beta = std::sqrt((height * width) / max_pixels);
+            h_bar = std::max(total_factor, floor_by_factor(height / beta));
+            w_bar = std::max(total_factor, floor_by_factor(width / beta));
+        } else if (h_bar * w_bar < min_pixels) {
+            const auto beta = std::sqrt(min_pixels / (height * width));
+            h_bar = ceil_by_factor(height * beta);
+            w_bar = ceil_by_factor(width * beta);
+        }
+
+        const std::array<uint8_t, 3> pad_color = {122, 116, 104};
+
+        clip_image_u8 resized_img;
+        image_manipulation::resize_and_pad_image(*img, resized_img, clip_image_size{w_bar, h_bar}, pad_color);
+        clip_image_f32_ptr res(clip_image_f32_init());
+        normalize_image_u8_to_f32(resized_img, *res, params.image_mean, params.image_std);
+        res_imgs->entries.push_back(std::move(res));
+        return true;
     }
 
     // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
@@ -3467,10 +3581,10 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         }
 
         return true;
-
+    } else {
+        GGML_ABORT("Unknown image preprocessing type");
     }
 
-    GGML_ASSERT(false && "Unknown image preprocessing type");
 }
 
 ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
@@ -3534,8 +3648,9 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 *
 int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
     const auto & params = ctx->model.hparams;
 
-    // only for models using fixed size square images
-    int n_patches_sq = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
+    // for models with fixed size image, the input image is already pre-processed and resized to square
+    int patch_size = params.patch_size;
+    int n_patches = (img->nx / patch_size) * (img->ny / patch_size);
 
     projector_type proj = ctx->proj_type();
 
@@ -3549,89 +3664,85 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_LDPV2:
         case PROJECTOR_TYPE_GLM_EDGE:
             {
-                n_patches_sq /= 4;
+                n_patches /= 4;
                 if (ctx->model.mm_glm_tok_boi) {
-                    n_patches_sq += 2; // for BOI and EOI token embeddings
+                    n_patches += 2; // for BOI and EOI token embeddings
                 }
             } break;
         case PROJECTOR_TYPE_MINICPMV:
             {
-                if (params.minicpmv_version == 2) {
-                    // MiniCPM-V 2.5
-                    n_patches_sq = 96;
-                } else if (params.minicpmv_version == 3) {
-                    // MiniCPM-V 2.6
-                    n_patches_sq = 64;
-                } else if (params.minicpmv_version == 4) {
-                    // MiniCPM-o 2.6
-                    n_patches_sq = 64;
-                } else if (params.minicpmv_version == 5) {
-                    // MiniCPM-V 4.0
-                    n_patches_sq = 64;
+                // Use actual config value if available, otherwise fall back to hardcoded values
+                if (params.minicpmv_query_num > 0) {
+                    n_patches = params.minicpmv_query_num;
                 } else {
-                    GGML_ABORT("Unknown minicpmv version");
+                    // Fallback to hardcoded values for legacy models
+                    if (params.minicpmv_version == 2) {
+                        n_patches = 96;
+                    } else if (params.minicpmv_version == 3) {
+                        n_patches = 64;
+                    } else if (params.minicpmv_version == 4) {
+                        n_patches = 64;
+                    } else if (params.minicpmv_version == 5) {
+                        // MiniCPM-V 4.0
+                        n_patches = 64;
+                    } else {
+                        GGML_ABORT("Unknown minicpmv version");
+                    }
                 }
             } break;
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:
             {
-                // dynamic size
+                // dynamic size (2 conv, so double patch size)
                 int patch_size = params.patch_size * 2;
                 int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0);
                 int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0);
-                n_patches_sq = x_patch * y_patch;
+                n_patches = x_patch * y_patch;
             } break;
         case PROJECTOR_TYPE_GEMMA3:
-            {
-                int n_per_side = params.image_size / params.patch_size;
-                int n_per_side_2d_pool = n_per_side / params.proj_scale_factor;
-                n_patches_sq = n_per_side_2d_pool * n_per_side_2d_pool;
-            } break;
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_INTERNVL:
+        case PROJECTOR_TYPE_LLAMA4:
+        case PROJECTOR_TYPE_LFM2:
             {
                 // both W and H are divided by proj_scale_factor
-                n_patches_sq /= (params.proj_scale_factor * params.proj_scale_factor);
+                int scale_factor = ctx->model.hparams.proj_scale_factor;
+                n_patches /= (scale_factor * scale_factor);
             } break;
         case PROJECTOR_TYPE_PIXTRAL:
             {
                 // dynamic size
                 int n_merge = params.spatial_merge_size;
-                int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1);
-                int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1);
-                n_patches_sq = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
-            } break;
-        case PROJECTOR_TYPE_LLAMA4:
-            {
-                int scale_factor = ctx->model.hparams.proj_scale_factor;
-                n_patches_sq /= (scale_factor * scale_factor);
+                int n_patches_x = img->nx / patch_size / (n_merge > 0 ? n_merge : 1);
+                int n_patches_y = img->ny / patch_size / (n_merge > 0 ? n_merge : 1);
+                n_patches = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
             } break;
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_ULTRAVOX:
         case PROJECTOR_TYPE_QWEN2A:
             {
-                n_patches_sq = img->nx;
+                n_patches = img->nx;
 
                 const int proj_stack_factor = ctx->model.hparams.proj_stack_factor;
                 if (ctx->model.audio_has_stack_frames()) {
                     GGML_ASSERT(proj_stack_factor > 0);
-                    const int n_len = CLIP_ALIGN(n_patches_sq, proj_stack_factor);
-                    n_patches_sq = n_len / proj_stack_factor;
+                    const int n_len = CLIP_ALIGN(n_patches, proj_stack_factor);
+                    n_patches = n_len / proj_stack_factor;
                 }
 
                 // whisper downscales input token by half after conv1d
-                n_patches_sq /= 2;
+                n_patches /= 2;
 
                 if (ctx->model.audio_has_avgpool()) {
                     // divide by 2 because of nn.AvgPool1d(2, stride=2)
-                    n_patches_sq /= 2;
+                    n_patches /= 2;
                 }
             } break;
         default:
             GGML_ABORT("unsupported projector type");
     }
 
-    return n_patches_sq;
+    return n_patches;
 }
 
 static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
@@ -4031,6 +4142,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_INTERNVL:
         case PROJECTOR_TYPE_QWEN2A:
         case PROJECTOR_TYPE_ULTRAVOX:
+        case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_VOXTRAL:
             {
                 // do nothing
@@ -4102,7 +4214,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
 }
 
 int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
-    const auto & hparams = ctx->model.hparams;
     switch (ctx->model.proj_type) {
         case PROJECTOR_TYPE_LDP:
             return ctx->model.mm_model_block_1_block_2_1_b->ne[0];
@@ -4114,20 +4225,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_MLP_NORM:
             return ctx->model.mm_3_b->ne[0];
         case PROJECTOR_TYPE_MINICPMV:
-            if (hparams.minicpmv_version == 2) {
-                // MiniCPM-V 2.5
-                return 4096;
-            } else if (hparams.minicpmv_version == 3) {
-                // MiniCPM-V 2.6
-                return 3584;
-            } else if (hparams.minicpmv_version == 4) {
-                // MiniCPM-o 2.6
-                return 3584;
-            } else if (hparams.minicpmv_version == 5) {
-                // MiniCPM-V 4.0
-                return 2560;
-            }
-            GGML_ABORT("Unknown minicpmv version");
+            return ctx->model.mm_model_proj->ne[0];
         case PROJECTOR_TYPE_GLM_EDGE:
             return ctx->model.mm_model_mlp_3_w->ne[1];
         case PROJECTOR_TYPE_QWEN2VL:
@@ -4146,6 +4244,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_model_proj->ne[1];
         case PROJECTOR_TYPE_QWEN2A:
             return ctx->model.mm_fc_w->ne[1];
+        case PROJECTOR_TYPE_LFM2:
+            return ctx->model.mm_2_w->ne[1];
         default:
             GGML_ABORT("Unknown projector type");
     }
diff --git a/tools/mtmd/clip.h b/tools/mtmd/clip.h
index 08f3efb7b1daf..3387cdbd36955 100644
--- a/tools/mtmd/clip.h
+++ b/tools/mtmd/clip.h
@@ -82,11 +82,6 @@ struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch
  */
 void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
 
-bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
-
-/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
-bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
-
 /** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
 bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
 
diff --git a/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py b/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py
index 3c60209544af3..4dda60a21164b 100644
--- a/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py
+++ b/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py
@@ -517,6 +517,16 @@ def bytes_to_unicode():
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
+# Read config.json to get actual model configuration
+config_path = os.path.join(dir_model, "config.json")
+model_config = {}
+if os.path.isfile(config_path):
+    with open(config_path, "r", encoding="utf-8") as f:
+        model_config = json.load(f)
+    print(f"Loaded config from {config_path}")
+else:
+    print(f"Warning: config.json not found at {config_path}")
+
 # If minicpmv_projector is not specified but the default path exists, use the default path
 if args.minicpmv_projector is None:
     default_projector_path = os.path.join(dir_model, "minicpmv.projector")
@@ -555,37 +565,62 @@ def bytes_to_unicode():
 #     processor = CLIPProcessor.from_pretrained(dir_model)
 
 minicpmv_version = args.minicpmv_version
-emb_dim = 4096
-block_count = 26
-if minicpmv_version == 1:  # MiniCPM-V 2.0
-    emb_dim = 2304
-    block_count = 26
-elif minicpmv_version == 2:  # MiniCPM-V 2.5
-    emb_dim = 4096
-    block_count = 27
-elif minicpmv_version == 3:  # MiniCPM-V 2.6
-    emb_dim = 3584
-    block_count = 27
-elif minicpmv_version == 4:  # MiniCPM-o 2.6
-    emb_dim = 3584
-    block_count = 27
-elif minicpmv_version == 5:  # MiniCPM-V 4.0
-    emb_dim = 2560
-    block_count = 27
-
-default_vision_config = {
-        "hidden_size": 1152,
-        "image_size": 980,
-        "intermediate_size": 4304,
-        "model_type": "idefics2",
-        "num_attention_heads": 16,
-        "num_hidden_layers": 27,
-        "patch_size": 14,
+
+# Use actual config values instead of hardcoded ones
+if model_config:
+    # For the projector/resampler, use the main model's hidden_size
+    emb_dim = model_config.get("hidden_size", 1536)
+
+    # For the vision model, use vision_config values
+    vision_config_dict = model_config.get("vision_config", {})
+    default_vision_config = {
+        "hidden_size": vision_config_dict.get("hidden_size", 1152),
+        "image_size": vision_config_dict.get("image_size", 980),
+        "intermediate_size": vision_config_dict.get("intermediate_size", 4304),
+        "model_type": vision_config_dict.get("model_type", "siglip"),
+        "num_attention_heads": vision_config_dict.get("num_attention_heads", 16),
+        "num_hidden_layers": vision_config_dict.get("num_hidden_layers", 27),
+        "patch_size": vision_config_dict.get("patch_size", 14),
     }
 
+    # Use vision model's num_hidden_layers for block_count
+    block_count = vision_config_dict.get("num_hidden_layers", 27)
+
+    print(f"Using config values: emb_dim={emb_dim}, block_count={block_count}")
+    print(f"Vision config: {default_vision_config}")
+else:
+    # Fallback to original hardcoded logic if config.json not found
+    emb_dim = 4096
+    block_count = 26
+    if minicpmv_version == 1:
+        emb_dim = 2304
+        block_count = 26
+    elif minicpmv_version == 2:
+        emb_dim = 4096
+        block_count = 27
+    elif minicpmv_version == 3:
+        emb_dim = 3584
+        block_count = 27
+    elif minicpmv_version == 4:
+        emb_dim = 3584
+        block_count = 27
+    elif minicpmv_version == 5:
+        emb_dim = 2560
+        block_count = 27
+
+    default_vision_config = {
+            "hidden_size": 1152,
+            "image_size": 980,
+            "intermediate_size": 4304,
+            "model_type": "idefics2",
+            "num_attention_heads": 16,
+            "num_hidden_layers": 27,
+            "patch_size": 14,
+        }
+
 vision_config = Idefics2VisionConfig(**default_vision_config)
 model = Idefics2VisionTransformer(vision_config)
-if minicpmv_version == 3:
+if minicpmv_version == 3 or (model_config and model_config.get("vision_config", {}).get("model_type") == "siglip"):
     vision_config = SiglipVisionConfig(**default_vision_config)
     model = SiglipVisionTransformer(vision_config)
 elif minicpmv_version == 4:
@@ -644,16 +679,27 @@ def bytes_to_unicode():
     fout.add_description("two-tower CLIP model")
 
 if has_vision_encoder:
-    # vision_model hparams
-    fout.add_uint32("clip.vision.image_size", 448)
-    fout.add_uint32("clip.vision.patch_size", 14)
-    fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), 1152)
-    fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), 4304)
+    # vision_model hparams - use actual config values
+    vision_image_size = model_config.get("image_size", 448) if model_config else 448
+    vision_patch_size = default_vision_config.get("patch_size", 14)
+    vision_hidden_size = default_vision_config.get("hidden_size", 1152)
+    vision_intermediate_size = default_vision_config.get("intermediate_size", 4304)
+    vision_attention_heads = default_vision_config.get("num_attention_heads", 16)
+
+    fout.add_uint32("clip.vision.image_size", vision_image_size)
+    fout.add_uint32("clip.vision.patch_size", vision_patch_size)
+    fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), vision_hidden_size)
+    fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), vision_intermediate_size)
     fout.add_uint32("clip.vision.projection_dim", 0)
-    fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
+    fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), vision_attention_heads)
     fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
     fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
 
+    # Add MiniCPM-V specific parameters
+    query_num = model_config.get("query_num", 0) if model_config else 0
+    resampler_emb_dim = model_config.get("hidden_size", 0) if model_config else 0
+    fout.add_uint32("clip.minicpmv_query_num", query_num)
+
     if processor is not None:
         image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
         image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
diff --git a/tools/mtmd/legacy-models/minicpmv-surgery.py b/tools/mtmd/legacy-models/minicpmv-surgery.py
index ba82116582b1f..53526623cd7cb 100644
--- a/tools/mtmd/legacy-models/minicpmv-surgery.py
+++ b/tools/mtmd/legacy-models/minicpmv-surgery.py
@@ -16,6 +16,8 @@
 
 # store these tensors in a new dictionary and torch.save them
 projector = {name: checkpoint[name].float() for name in mm_tensors}
+if 'resampler.proj' in projector.keys() and hasattr(model.llm.config,'scale_emb') is True:
+    projector['resampler.proj'] = projector['resampler.proj'] / model.llm.config.scale_emb
 torch.save(projector, f"{args.model}/minicpmv.projector")
 
 clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
diff --git a/tools/mtmd/mtmd-cli.cpp b/tools/mtmd/mtmd-cli.cpp
index 599e682e0f894..5fde6ca0c32ae 100644
--- a/tools/mtmd/mtmd-cli.cpp
+++ b/tools/mtmd/mtmd-cli.cpp
@@ -108,7 +108,7 @@ struct mtmd_cli_context {
         }
 
         tmpls = common_chat_templates_init(model, params.chat_template);
-        LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja).c_str());
+        LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja, params.default_template_kwargs).c_str());
 
         init_vision_context(params);
 
diff --git a/tools/mtmd/tests.sh b/tools/mtmd/tests.sh
index e73cf96af2941..6f8a5f86ac5b2 100755
--- a/tools/mtmd/tests.sh
+++ b/tools/mtmd/tests.sh
@@ -68,6 +68,7 @@ add_test_vision "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
 add_test_vision "ggml-org/InternVL2_5-1B-GGUF:Q8_0"
 add_test_vision "ggml-org/InternVL3-1B-Instruct-GGUF:Q8_0"
 add_test_vision "ggml-org/Qwen2.5-Omni-3B-GGUF:Q4_K_M"
+add_test_vision "ggml-org/LFM2-VL-450M-GGUF:Q8_0"
 
 add_test_audio  "ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF:Q8_0"
 add_test_audio  "ggml-org/Qwen2.5-Omni-3B-GGUF:Q4_K_M"
diff --git a/tools/perplexity/perplexity.cpp b/tools/perplexity/perplexity.cpp
index 189dcb3d72f5e..80cbb095da4cb 100644
--- a/tools/perplexity/perplexity.cpp
+++ b/tools/perplexity/perplexity.cpp
@@ -525,7 +525,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
     }
 
     // We get the logits for all the tokens in the context window (params.n_ctx)
-    // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
+    // from llama_decode below.  Now, based on https://huggingface.co/docs/transformers/perplexity,
     // calculate the perplexity over the last half of the window (so the model always has
     // some context to predict the token).
     //
@@ -559,7 +559,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
             for (int seq = 0; seq < n_seq_batch; seq++) {
                 int seq_start = batch_start + seq*n_ctx;
 
-                // save original token and restore it after eval
+                // save original token and restore it after decode
                 const auto token_org = tokens[seq_start];
 
                 // add BOS token for the first batch of each chunk
@@ -584,7 +584,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
             }
 
             if (llama_decode(ctx, batch)) {
-                LOG_INF("%s : failed to eval\n", __func__);
+                LOG_INF("%s : failed to decode\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
             }
 
@@ -920,7 +920,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
         }
 
         if (i0 == i1) {
-            LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
+            LOG_ERR("%s : task %zu does not fit in the context window (requires %lu tokens)\n", __func__, i0, hs_data[i0].required_tokens);
             return;
         }
 
@@ -1213,7 +1213,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
         }
 
         if (i0 == i1) {
-            LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
+            LOG_ERR("%s : task %zu does not fit in the context window (requires %lu tokens)\n", __func__, i0, data[i0].required_tokens);
             return;
         }
 
@@ -1548,6 +1548,10 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
 
             int num_answers = cur_task.seq_tokens.size();
             if (s0 + num_answers > max_seq) {
+                if (s0 == 0) {
+                    LOG_ERR("%s : task %zu requires a higher -np|--parallel value (at least %d)\n", __func__, i0, num_answers);
+                    return;
+                }
                 break;
             }
 
@@ -1588,7 +1592,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
         }
 
         if (i0 == i1) {
-            LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
+            LOG_ERR("%s : task %zu does not fit in the context window (requires %lu tokens)\n", __func__, i0, tasks[i0].required_tokens);
             return;
         }
 
diff --git a/tools/server/CMakeLists.txt b/tools/server/CMakeLists.txt
index c2a56aaa7f27b..83b608c32a9c7 100644
--- a/tools/server/CMakeLists.txt
+++ b/tools/server/CMakeLists.txt
@@ -33,7 +33,7 @@ endforeach()
 add_executable(${TARGET} ${TARGET_SRCS})
 install(TARGETS ${TARGET} RUNTIME)
 
-target_include_directories(${TARGET} PRIVATE ../llava)
+target_include_directories(${TARGET} PRIVATE ../mtmd)
 target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
 target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
 
diff --git a/tools/server/README.md b/tools/server/README.md
index 87cef75730afb..baf3730add67c 100644
--- a/tools/server/README.md
+++ b/tools/server/README.md
@@ -66,7 +66,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
 | `-ctk, --cache-type-k TYPE` | KV cache data type for K<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
 | `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
-| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: 0.1, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
+| `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_NO_MMAP) |
@@ -226,6 +226,10 @@ services:
 ### Multimodal support
 
 Multimodal support was added in [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) and is currently an experimental feature.
+It is currently available in the following endpoints:
+- The OAI-compatible chat endpoint.
+- The non-OAI-compatible completions endpoint.
+- The non-OAI-compatible embeddings endpoint.
 
 For more details, please refer to [multimodal documentation](../../docs/multimodal.md)
 
@@ -400,12 +404,15 @@ These input shapes and data type are allowed for `prompt`:
   - Single string: `"string"`
   - Single sequence of tokens: `[12, 34, 56]`
   - Mixed tokens and strings: `[12, 34, "string", 56, 78]`
+  - A JSON object which optionally contains multimodal data: `{ "prompt_string": "string", "multimodal_data": ["base64"] }`
 
 Multiple prompts are also supported. In this case, the completion result will be an array.
 
   - Only strings: `["string1", "string2"]`
-  - Strings and sequences of tokens: `["string1", [12, 34, 56]]`
-  - Mixed types: `[[12, 34, "string", 56, 78], [12, 34, 56], "string"]`
+  - Strings, JSON objects, and sequences of tokens: `["string1", [12, 34, 56], { "prompt_string": "string", "multimodal_data": ["base64"]}]`
+  - Mixed types: `[[12, 34, "string", 56, 78], [12, 34, 56], "string", { "prompt_string": "string" }]`
+
+Note for `multimodal_data` in JSON object prompts. This should be an array of strings, containing base64 encoded multimodal data such as images and audio. There must be an identical number of MTMD media markers in the string prompt element which act as placeholders for the data provided to this parameter. The multimodal data files will be substituted in order. The marker string (e.g. `<__media__>`) can be found by calling `mtmd_default_marker()` defined in [the MTMD C API](https://github.com/ggml-org/llama.cpp/blob/5fd160bbd9d70b94b5b11b0001fd7f477005e4a0/tools/mtmd/mtmd.h#L87). A client *must not* specify this field unless the server has the multimodal capability. Clients should check `/models` or `/v1/models` for the `multimodal` capability before a multimodal request.
 
 `temperature`: Adjust the randomness of the generated text. Default: `0.8`
 
@@ -477,8 +484,6 @@ These words will not be included in the completion, so make sure to add them to
 
 `t_max_predict_ms`: Set a time limit in milliseconds for the prediction (a.k.a. text-generation) phase. The timeout will trigger if the generation takes more than the specified time (measured since the first token was generated) and if a new-line character has already been generated. Useful for FIM applications. Default: `0`, which is disabled.
 
-`image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
-
 `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot.  Default: `-1`
 
 `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `true`
@@ -638,12 +643,12 @@ Returns a JSON object with a field `prompt` containing a string of the input mes
 
 The same as [the embedding example](../embedding) does.
 
+This endpoint also supports multimodal embeddings. See the documentation for the `/completions` endpoint for details on how to send a multimodal prompt.
+
 *Options:*
 
 `content`: Set the text to process.
 
-`image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
-
 `embd_normalize`: Normalization for pooled embeddings. Can be one of the following values:
 ```
   -1: No normalization
@@ -1132,6 +1137,12 @@ The `response_format` parameter supports both plain JSON output (e.g. `{"type":
 
 `chat_template_kwargs`: Allows sending additional parameters to the json templating system. For example: `{"enable_thinking": false}`
 
+`reasoning_format`: The reasoning format to be parsed. If set to `none`, it will output the raw generated text.
+
+`thinking_forced_open`: Force a reasoning model to always output the reasoning. Only works on certain models.
+
+`parse_tool_calls`: Whether to parse the generated tool call.
+
 *Examples:*
 
 You can use either Python `openai` library with appropriate checkpoints:
diff --git a/tools/server/bench/bench.py b/tools/server/bench/bench.py
index 5cc6f92ab6c53..0c57a2df04a60 100644
--- a/tools/server/bench/bench.py
+++ b/tools/server/bench/bench.py
@@ -274,7 +274,6 @@ def start_server_background(args):
     server_args.extend(['--batch-size', args.batch_size])
     server_args.extend(['--ubatch-size', args.ubatch_size])
     server_args.extend(['--n-predict', args.max_tokens * 2])
-    server_args.extend(['--defrag-thold', "0.1"])
     server_args.append('--cont-batching')
     server_args.append('--metrics')
     server_args.append('--flash-attn')
diff --git a/tools/server/public/index.html.gz b/tools/server/public/index.html.gz
index 4e25eef1760ca..0e3e1f9834a7f 100644
Binary files a/tools/server/public/index.html.gz and b/tools/server/public/index.html.gz differ
diff --git a/tools/server/server.cpp b/tools/server/server.cpp
index a255d481a4d1c..6eb5aeb582b3a 100644
--- a/tools/server/server.cpp
+++ b/tools/server/server.cpp
@@ -383,8 +383,12 @@ struct server_task {
             } else {
                 params.oaicompat_chat_syntax.format = defaults.oaicompat_chat_syntax.format;
             }
-            params.oaicompat_chat_syntax.reasoning_format = params_base.reasoning_format;
-            params.oaicompat_chat_syntax.reasoning_in_content = params.stream && (params_base.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY);
+            common_reasoning_format reasoning_format = params_base.reasoning_format;
+            if (data.contains("reasoning_format")) {
+                reasoning_format = common_reasoning_format_from_name(data.at("reasoning_format").get<std::string>());
+            }
+            params.oaicompat_chat_syntax.reasoning_format = reasoning_format;
+            params.oaicompat_chat_syntax.reasoning_in_content = params.stream && (reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY);
             params.oaicompat_chat_syntax.thinking_forced_open = json_value(data, "thinking_forced_open", false);
             params.oaicompat_chat_syntax.parse_tool_calls = json_value(data, "parse_tool_calls", false);
         }
@@ -688,6 +692,13 @@ struct completion_token_output {
     }
 };
 
+struct swa_checkpoint {
+    llama_pos pos_min;
+    llama_pos pos_max;
+
+    std::vector<uint8_t> data;
+};
+
 struct server_task_result_cmpl_final : server_task_result {
     int index = 0;
 
@@ -900,6 +911,17 @@ struct server_task_result_cmpl_final : server_task_result {
             {"model",              oaicompat_model},
             {"system_fingerprint", build_info},
             {"object",             "chat.completion.chunk"},
+        });
+
+        // OpenAI API spec for chat.completion.chunks specifies an empty `choices` array for the last chunk when including usage
+        // https://platform.openai.com/docs/api-reference/chat_streaming/streaming#chat_streaming/streaming-choices
+        deltas.push_back({
+            {"choices", json::array()},
+            {"created",            t},
+            {"id",                 oaicompat_cmpl_id},
+            {"model",              oaicompat_model},
+            {"system_fingerprint", build_info},
+            {"object",             "chat.completion.chunk"},
             {"usage", json {
                 {"completion_tokens", n_decoded},
                 {"prompt_tokens",     n_prompt_tokens},
@@ -1190,6 +1212,8 @@ struct server_task_result_metrics : server_task_result {
     uint64_t n_tokens_predicted_total        = 0;
     uint64_t t_tokens_generation_total       = 0;
 
+    uint64_t n_past_max = 0;
+
     uint64_t n_prompt_tokens_processed = 0;
     uint64_t t_prompt_processing       = 0;
 
@@ -1215,6 +1239,8 @@ struct server_task_result_metrics : server_task_result {
             { "n_tokens_predicted_total",        n_tokens_predicted_total },
             { "t_prompt_processing_total",       t_prompt_processing_total },
 
+            { "n_past_max",                      n_past_max },
+
             { "n_prompt_tokens_processed",       n_prompt_tokens_processed },
             { "t_prompt_processing",             t_prompt_processing },
             { "n_tokens_predicted",              n_tokens_predicted },
@@ -1332,6 +1358,8 @@ struct server_slot {
 
     std::vector<completion_token_output> generated_token_probs;
 
+    std::vector<swa_checkpoint> swa_checkpoints;
+
     bool has_next_token = true;
     bool has_new_line   = false;
     bool truncated      = false;
@@ -1574,6 +1602,8 @@ struct server_metrics {
     uint64_t n_tokens_predicted_total        = 0;
     uint64_t t_tokens_generation_total       = 0;
 
+    uint64_t n_past_max = 0;
+
     uint64_t n_prompt_tokens_processed = 0;
     uint64_t t_prompt_processing       = 0;
 
@@ -1592,6 +1622,10 @@ struct server_metrics {
         n_prompt_tokens_processed       += slot.n_prompt_tokens_processed;
         t_prompt_processing             += slot.t_prompt_processing;
         t_prompt_processing_total       += slot.t_prompt_processing;
+
+        if (slot.n_past > 0) {
+            n_past_max = std::max(n_past_max, (uint64_t) slot.n_past);
+        }
     }
 
     void on_prediction(const server_slot & slot) {
@@ -1607,6 +1641,9 @@ struct server_metrics {
             if (slot.is_processing()) {
                 n_busy_slots_total++;
             }
+            if (slot.n_past > 0) {
+                n_past_max = std::max(n_past_max, (uint64_t) slot.n_past);
+            }
         }
     }
 
@@ -1703,7 +1740,7 @@ struct server_queue {
     void pop_deferred_task() {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         if (!queue_tasks_deferred.empty()) {
-            queue_tasks.emplace_back(std::move(queue_tasks_deferred.front()));
+            queue_tasks.emplace_front(std::move(queue_tasks_deferred.front()));
             queue_tasks_deferred.pop_front();
         }
         condition_tasks.notify_one();
@@ -2011,6 +2048,10 @@ struct server_context {
             params_dft.cache_type_k = params_base.speculative.cache_type_k;
             params_dft.cache_type_v = params_base.speculative.cache_type_v;
 
+            params_dft.cpuparams.n_threads = params_base.speculative.cpuparams.n_threads;
+            params_dft.cpuparams_batch.n_threads = params_base.speculative.cpuparams_batch.n_threads;
+            params_dft.tensor_buft_overrides = params_base.speculative.tensor_buft_overrides;
+
             llama_init_dft = common_init_from_params(params_dft);
 
             model_dft = llama_init_dft.model.get();
@@ -2036,7 +2077,7 @@ struct server_context {
 
         chat_templates = common_chat_templates_init(model, params_base.chat_template);
         try {
-            common_chat_format_example(chat_templates.get(), params.use_jinja);
+            common_chat_format_example(chat_templates.get(), params.use_jinja, params.default_template_kwargs);
         } catch (const std::exception & e) {
             SRV_WRN("%s: Chat template parsing error: %s\n", __func__, e.what());
             SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
@@ -2858,6 +2899,8 @@ struct server_context {
                     res->n_tokens_predicted_total        = metrics.n_tokens_predicted_total;
                     res->t_tokens_generation_total       = metrics.t_tokens_generation_total;
 
+                    res->n_past_max = metrics.n_past_max;
+
                     res->n_prompt_tokens_processed = metrics.n_prompt_tokens_processed;
                     res->t_prompt_processing       = metrics.t_prompt_processing;
                     res->n_tokens_predicted        = metrics.n_tokens_predicted;
@@ -3285,6 +3328,8 @@ struct server_context {
                                 slot.n_past = 0;
                             }
 
+                            const auto n_swa = llama_model_n_swa(model);
+
                             if (slot.n_past > 0 && slot.n_past < (int) slot.cache_tokens.size()) {
                                 const auto pos_min = llama_memory_seq_pos_min(llama_get_memory(ctx), slot.id);
                                 if (pos_min == -1) {
@@ -3292,12 +3337,58 @@ struct server_context {
                                     GGML_ABORT("pos_min == -1, but n_past > 0 - should not happen: https://github.com/ggml-org/llama.cpp/pull/13833#discussion_r2116181237");
                                 }
 
-                                const auto n_swa = llama_model_n_swa(model);
-                                if (pos_min > std::max(0, slot.n_past - n_swa)) {
+                                const auto pos_min_thold = std::max(0, slot.n_past - n_swa);
+
+                                if (pos_min > pos_min_thold) {
                                     SLT_WRN(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d, n_swa = %d\n", slot.n_past, (int) slot.cache_tokens.size(), slot.id, pos_min, n_swa);
-                                    SLT_WRN(slot, "forcing full prompt re-processing due to lack of cache data (likely due to SWA, see %s)\n",
-                                            "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
-                                    slot.n_past = 0;
+
+                                    // search for a SWA checkpoint
+                                    const auto it = std::find_if(
+                                        slot.swa_checkpoints.rbegin(),
+                                        slot.swa_checkpoints.rend(),
+                                        [&](const auto & cur) {
+                                            return cur.pos_min <= pos_min_thold;
+                                        }
+                                    );
+
+                                    bool do_reset = it == slot.swa_checkpoints.rend();
+
+                                    if (!do_reset) {
+                                        // restore the checkpoint
+                                        const size_t swa_size = it->data.size();
+                                        const size_t n = llama_state_seq_set_data_ext(ctx, it->data.data(), swa_size, slot.id, LLAMA_STATE_SEQ_FLAGS_SWA_ONLY);
+
+                                        if (n != swa_size) {
+                                            SLT_ERR(slot, "failed to restore SWA checkpoint, pos_min = %d, pos_max = %d, size = %.3f MiB\n", it->pos_min, it->pos_max, (float) swa_size / 1024 / 1024);
+                                            do_reset = true;
+                                        } else {
+                                            slot.n_past = std::min(slot.n_past, it->pos_max);
+
+                                            SLT_WRN(slot, "SWA checkpoint restore, pos_min = %d, pos_max = %d, size = %.3f MiB\n", it->pos_min, it->pos_max, (float) swa_size / 1024 / 1024);
+                                        }
+                                    }
+
+                                    if (do_reset) {
+                                        SLT_WRN(slot, "forcing full prompt re-processing due to lack of cache data (likely due to SWA, see %s)\n",
+                                                "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
+
+                                        slot.n_past = 0;
+                                        slot.swa_checkpoints.clear();
+                                    }
+                                }
+                            }
+
+                            if (n_swa > 0) {
+                                const auto pos_min_thold = std::max(0, slot.n_past - n_swa);
+
+                                // erase any checkpoints with pos_min > pos_min_thold
+                                for (int i = (int) slot.swa_checkpoints.size() - 1; i >= 0; i--) {
+                                    const auto & cur = slot.swa_checkpoints[i];
+                                    if (cur.pos_min > pos_min_thold) {
+                                        slot.swa_checkpoints.erase(slot.swa_checkpoints.begin() + i);
+
+                                        SLT_WRN(slot, "SWA checkpoint erase, pos_min = %d, pos_max = %d, size = %.3f MiB\n", cur.pos_min, cur.pos_max, (float) cur.data.size() / 1024 / 1024);
+                                    }
                                 }
                             }
                         }
@@ -3511,6 +3602,39 @@ struct server_context {
 
                     // prompt evaluated for next-token prediction
                     slot.state = SLOT_STATE_GENERATING;
+
+                    // make a checkpoint with the SWA memory
+                    // checkpoints are needed only if we are not using "--swa-full"
+                    if (llama_model_n_swa(model) > 0 && !params_base.swa_full && params_base.n_swa_checkpoints > 0) {
+                        if (slot.swa_checkpoints.size() >= (size_t) params_base.n_swa_checkpoints) {
+                            {
+                                const auto & cur = slot.swa_checkpoints.back();
+
+                                SLT_WRN(slot, "SWA checkpoint erase, pos_min = %d, pos_max = %d, size = %.3f MiB\n",
+                                        cur.pos_min, cur.pos_max, (float) cur.data.size() / 1024 / 1024);
+                            }
+
+                            slot.swa_checkpoints.erase(slot.swa_checkpoints.begin());
+                        }
+
+                        const size_t swa_size = llama_state_seq_get_size_ext(ctx, slot.id, LLAMA_STATE_SEQ_FLAGS_SWA_ONLY);
+
+                        auto & cur = slot.swa_checkpoints.emplace_back(swa_checkpoint{
+                            /*.pos_min = */ llama_memory_seq_pos_min(llama_get_memory(ctx), slot.id),
+                            /*.pos_max = */ llama_memory_seq_pos_max(llama_get_memory(ctx), slot.id),
+                            /*.data    = */ std::vector<uint8_t>(swa_size),
+                        });
+
+                        llama_state_seq_get_data_ext(ctx, cur.data.data(), swa_size, slot.id, LLAMA_STATE_SEQ_FLAGS_SWA_ONLY);
+
+                        float size_total = 0.0f;
+                        for (const auto & checkpoint : slot.swa_checkpoints) {
+                            size_total += (float) checkpoint.data.size() / 1024 / 1024;
+                        }
+
+                        SLT_WRN(slot, "SWA checkpoint create, pos_min = %d, pos_max = %d, size = %.3f MiB, total = %d/%d (%.3f MiB)\n",
+                                cur.pos_min, cur.pos_max, (float) cur.data.size() / 1024 / 1024, (int) slot.swa_checkpoints.size(), params_base.n_swa_checkpoints, size_total);
+                    }
                 } else if (slot.state != SLOT_STATE_GENERATING) {
                     continue; // continue loop of slots
                 }
@@ -3979,6 +4103,10 @@ int main(int argc, char ** argv) {
                     {"name",  "n_decode_total"},
                     {"help",  "Total number of llama_decode() calls"},
                     {"value",  res_metrics->n_decode_total}
+            }, {
+                    {"name",  "n_past_max"},
+                    {"help",  "Largest observed n_past."},
+                    {"value",  res_metrics->n_past_max}
             }, {
                     {"name",  "n_busy_slots_per_decode"},
                     {"help",  "Average number of busy slots per llama_decode() call"},
@@ -4181,6 +4309,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_api_show = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
+        bool has_mtmd = ctx_server.mctx != nullptr;
         json data = {
             {
                 "template", common_chat_templates_source(ctx_server.chat_templates.get()),
@@ -4202,7 +4331,7 @@ int main(int argc, char ** argv) {
                 {"quantization_level", ""}
             }},
             {"model_info", ""},
-            {"capabilities", {"completion"}}
+            {"capabilities", has_mtmd ? json({"completion","multimodal"}) : json({"completion"})}
         };
 
         res_ok(res, data);
@@ -4228,56 +4357,15 @@ int main(int argc, char ** argv) {
             // TODO: this log can become very long, put it behind a flag or think about a more compact format
             //SRV_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get<std::string>().c_str() : prompt.dump(2).c_str());
 
-            // process files
-            mtmd::bitmaps bitmaps;
-            const bool has_mtmd = ctx_server.mctx != nullptr;
-            {
-                if (!has_mtmd && !files.empty()) {
-                    throw std::runtime_error("This server does not support multimodal");
-                }
-                for (auto & file : files) {
-                    mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(ctx_server.mctx, file.data(), file.size()));
-                    if (!bmp.ptr) {
-                        throw std::runtime_error("Failed to load image or audio file");
-                    }
-                    // calculate bitmap hash (for KV caching)
-                    std::string hash = fnv_hash(bmp.data(), bmp.n_bytes());
-                    bmp.set_id(hash.c_str());
-                    bitmaps.entries.push_back(std::move(bmp));
-                }
-            }
-
             // process prompt
             std::vector<server_tokens> inputs;
 
-            if (oaicompat && has_mtmd) {
-                // multimodal
-                std::string prompt_str = prompt.get<std::string>();
-                mtmd_input_text inp_txt = {
-                    prompt_str.c_str(),
-                    /* add_special */   true,
-                    /* parse_special */ true,
-                };
-                mtmd::input_chunks chunks(mtmd_input_chunks_init());
-                auto bitmaps_c_ptr = bitmaps.c_ptr();
-                int32_t tokenized = mtmd_tokenize(ctx_server.mctx,
-                                                    chunks.ptr.get(),
-                                                    &inp_txt,
-                                                    bitmaps_c_ptr.data(),
-                                                    bitmaps_c_ptr.size());
-                if (tokenized != 0) {
-                    throw std::runtime_error("Failed to tokenize prompt");
-                }
-
-                server_tokens tmp(chunks, true);
-                inputs.push_back(std::move(tmp));
+            if (oaicompat && ctx_server.mctx != nullptr) {
+                // This is the case used by OAI compatible chat path with MTMD. TODO It can be moved to the path below.
+                inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get<std::string>(), files));
             } else {
-                // non-multimodal version
-                auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
-                for (auto & p : tokenized_prompts) {
-                    auto tmp = server_tokens(p, ctx_server.mctx != nullptr);
-                    inputs.push_back(std::move(tmp));
-                }
+                // Everything else, including multimodal completions.
+                inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
             }
 
             tasks.reserve(inputs.size());
@@ -4446,7 +4534,7 @@ int main(int argc, char ** argv) {
         data["input_extra"] = input_extra; // default to empty array if it's not exist
 
         std::string prompt = json_value(data, "prompt", std::string());
-        std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, false, true);
+        std::vector<server_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, false, true);
         SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size());
         data["prompt"] = format_infill(
             ctx_server.vocab,
@@ -4457,7 +4545,7 @@ int main(int argc, char ** argv) {
             ctx_server.params_base.n_predict,
             ctx_server.slots[0].n_ctx, // TODO: there should be a better way
             ctx_server.params_base.spm_infill,
-            tokenized_prompts[0]
+            tokenized_prompts[0].get_text_tokens() // TODO: this could maybe be multimodal.
         );
 
         std::vector<raw_buffer> files; // dummy
@@ -4506,7 +4594,7 @@ int main(int argc, char ** argv) {
         if (current_state == SERVER_STATE_READY) {
             model_meta = ctx_server.model_meta();
         }
-
+        bool has_mtmd = ctx_server.mctx != nullptr;
         json models = {
             {"models", {
                 {
@@ -4518,7 +4606,7 @@ int main(int argc, char ** argv) {
                     {"type", "model"},
                     {"description", ""},
                     {"tags", {""}},
-                    {"capabilities", {"completion"}},
+                    {"capabilities", has_mtmd ? json({"completion","multimodal"}) : json({"completion"})},
                     {"parameters", ""},
                     {"details", {
                         {"parent_model", ""},
@@ -4635,7 +4723,7 @@ int main(int argc, char ** argv) {
             }
         }
 
-        auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
+        auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
         for (const auto & tokens : tokenized_prompts) {
             // this check is necessary for models that do not add BOS token to the input
             if (tokens.empty()) {
@@ -4663,7 +4751,7 @@ int main(int argc, char ** argv) {
 
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
-                task.prompt_tokens = server_tokens(tokenized_prompts[i], ctx_server.mctx != nullptr);
+                task.prompt_tokens = std::move(tokenized_prompts[i]);
 
                 // OAI-compat
                 task.params.oaicompat = oaicompat;
@@ -4750,7 +4838,10 @@ int main(int argc, char ** argv) {
             return;
         }
 
-        llama_tokens tokenized_query = tokenize_input_prompts(ctx_server.vocab, query, /* add_special */ false, true)[0];
+        std::vector<server_tokens> tokenized_queries = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, query, /* add_special */ false, true);
+        if (tokenized_queries.size() != 1) {
+            res_error(res, format_error_response("\"query\" must contain only a single prompt", ERROR_TYPE_INVALID_REQUEST));
+        }
 
         // create and queue the task
         json responses = json::array();
@@ -4758,14 +4849,14 @@ int main(int argc, char ** argv) {
         std::unordered_set<int> task_ids;
         {
             std::vector<server_task> tasks;
-            auto tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true);
+            auto tokenized_docs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, documents, /* add_special */ false, true);
             tasks.reserve(tokenized_docs.size());
             for (size_t i = 0; i < tokenized_docs.size(); i++) {
-                auto tmp = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]);
+                auto tmp = format_rerank(ctx_server.vocab, tokenized_queries[0], tokenized_docs[i]);
                 server_task task   = server_task(SERVER_TASK_TYPE_RERANK);
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
-                task.prompt_tokens = server_tokens(tmp, ctx_server.mctx != nullptr);
+                task.prompt_tokens = std::move(tmp);
                 tasks.push_back(std::move(task));
             }
 
@@ -4977,7 +5068,7 @@ int main(int argc, char ** argv) {
     // print sample chat example to make it clear which template is used
     LOG_INF("%s: chat template, chat_template: %s, example_format: '%s'\n", __func__,
         common_chat_templates_source(ctx_server.chat_templates.get()),
-        common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja).c_str());
+        common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja, ctx_server.params_base.default_template_kwargs).c_str());
 
     ctx_server.queue_tasks.on_new_task([&ctx_server](server_task && task) {
         ctx_server.process_single_task(std::move(task));
diff --git a/tools/server/tests/unit/test_basic.py b/tools/server/tests/unit/test_basic.py
index 1485de8ceb3fc..c7b3af0489164 100644
--- a/tools/server/tests/unit/test_basic.py
+++ b/tools/server/tests/unit/test_basic.py
@@ -5,7 +5,7 @@
 server = ServerPreset.tinyllama2()
 
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinyllama2()
diff --git a/tools/server/tests/unit/test_chat_completion.py b/tools/server/tests/unit/test_chat_completion.py
index 6c6f64f5e2ec4..509c024b75fd3 100644
--- a/tools/server/tests/unit/test_chat_completion.py
+++ b/tools/server/tests/unit/test_chat_completion.py
@@ -72,27 +72,29 @@ def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_conte
     content = ""
     last_cmpl_id = None
     for i, data in enumerate(res):
-        choice = data["choices"][0]
-        if i == 0:
-            # Check first role message for stream=True
-            assert choice["delta"]["content"] is None
-            assert choice["delta"]["role"] == "assistant"
+        if data["choices"]:
+            choice = data["choices"][0]
+            if i == 0:
+                # Check first role message for stream=True
+                assert choice["delta"]["content"] is None
+                assert choice["delta"]["role"] == "assistant"
+            else:
+                assert "role" not in choice["delta"]
+            assert data["system_fingerprint"].startswith("b")
+            assert "gpt-3.5" in data["model"] # DEFAULT_OAICOMPAT_MODEL, maybe changed in the future
+            if last_cmpl_id is None:
+                last_cmpl_id = data["id"]
+            assert last_cmpl_id == data["id"] # make sure the completion id is the same for all events in the stream
+            if choice["finish_reason"] in ["stop", "length"]:
+                assert "content" not in choice["delta"]
+                assert match_regex(re_content, content)
+                assert choice["finish_reason"] == finish_reason
+            else:
+                assert choice["finish_reason"] is None
+                content += choice["delta"]["content"] or ''
         else:
-            assert "role" not in choice["delta"]
-        assert data["system_fingerprint"].startswith("b")
-        assert "gpt-3.5" in data["model"] # DEFAULT_OAICOMPAT_MODEL, maybe changed in the future
-        if last_cmpl_id is None:
-            last_cmpl_id = data["id"]
-        assert last_cmpl_id == data["id"] # make sure the completion id is the same for all events in the stream
-        if choice["finish_reason"] in ["stop", "length"]:
             assert data["usage"]["prompt_tokens"] == n_prompt
             assert data["usage"]["completion_tokens"] == n_predicted
-            assert "content" not in choice["delta"]
-            assert match_regex(re_content, content)
-            assert choice["finish_reason"] == finish_reason
-        else:
-            assert choice["finish_reason"] is None
-            content += choice["delta"]["content"] or ''
 
 
 def test_chat_completion_with_openai_library():
@@ -278,12 +280,14 @@ def test_chat_completion_with_timings_per_token():
             assert data["choices"][0]["delta"]["role"] == "assistant"
             assert "timings" not in data, f'First event should not have timings: {data}'
         else:
-            assert "role" not in data["choices"][0]["delta"]
-            assert "timings" in data
-            assert "prompt_per_second" in data["timings"]
-            assert "predicted_per_second" in data["timings"]
-            assert "predicted_n" in data["timings"]
-            assert data["timings"]["predicted_n"] <= 10
+            if data["choices"]:
+                assert "role" not in data["choices"][0]["delta"]
+            else:
+                assert "timings" in data
+                assert "prompt_per_second" in data["timings"]
+                assert "predicted_per_second" in data["timings"]
+                assert "predicted_n" in data["timings"]
+                assert data["timings"]["predicted_n"] <= 10
 
 
 def test_logprobs():
@@ -332,24 +336,25 @@ def test_logprobs_stream():
     output_text = ''
     aggregated_text = ''
     for i, data in enumerate(res):
-        choice = data.choices[0]
-        if i == 0:
-            # Check first role message for stream=True
-            assert choice.delta.content is None
-            assert choice.delta.role == "assistant"
-        else:
-            assert choice.delta.role is None
-            if choice.finish_reason is None:
-                if choice.delta.content:
-                    output_text += choice.delta.content
-                assert choice.logprobs is not None
-                assert choice.logprobs.content is not None
-                for token in choice.logprobs.content:
-                    aggregated_text += token.token
-                    assert token.logprob <= 0.0
-                    assert token.bytes is not None
-                    assert token.top_logprobs is not None
-                    assert len(token.top_logprobs) > 0
+        if data.choices:
+            choice = data.choices[0]
+            if i == 0:
+                # Check first role message for stream=True
+                assert choice.delta.content is None
+                assert choice.delta.role == "assistant"
+            else:
+                assert choice.delta.role is None
+                if choice.finish_reason is None:
+                    if choice.delta.content:
+                        output_text += choice.delta.content
+                    assert choice.logprobs is not None
+                    assert choice.logprobs.content is not None
+                    for token in choice.logprobs.content:
+                        aggregated_text += token.token
+                        assert token.logprob <= 0.0
+                        assert token.bytes is not None
+                        assert token.top_logprobs is not None
+                        assert len(token.top_logprobs) > 0
     assert aggregated_text == output_text
 
 
diff --git a/tools/server/tests/unit/test_completion.py b/tools/server/tests/unit/test_completion.py
index be3a0052c64fe..11483e679a505 100644
--- a/tools/server/tests/unit/test_completion.py
+++ b/tools/server/tests/unit/test_completion.py
@@ -6,8 +6,10 @@
 
 server = ServerPreset.tinyllama2()
 
+JSON_MULTIMODAL_KEY = "multimodal_data"
+JSON_PROMPT_STRING_KEY = "prompt_string"
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinyllama2()
@@ -229,8 +231,30 @@ def test_nocache_long_input_prompt():
         "temperature": 1.0,
         "cache_prompt": False,
     })
+    assert res.status_code == 400
+
+def test_json_prompt_no_mtmd():
+    global server
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "prompt": { JSON_PROMPT_STRING_KEY: "I believe the meaning of life is" },
+        "seed": 42,
+        "temperature": 1.0,
+        "cache_prompt": False,
+    })
     assert res.status_code == 200
 
+def test_json_prompt_mtm_error_when_not_supported():
+    global server
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "prompt": { JSON_PROMPT_STRING_KEY: "I believe the meaning of life is <__media__>", JSON_MULTIMODAL_KEY: "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=" },
+        "seed": 42,
+        "temperature": 1.0,
+        "cache_prompt": False,
+    })
+    # MTMD is disabled on this model, so this should fail.
+    assert res.status_code != 200
 
 def test_completion_with_tokens_input():
     global server
@@ -269,6 +293,20 @@ def test_completion_with_tokens_input():
     assert len(res.body) == 2
     assert res.body[0]["content"] == res.body[1]["content"]
 
+    # mixed JSON and tokens
+    res = server.make_request("POST", "/completion", data={
+        "prompt": [
+            tokens,
+            {
+                JSON_PROMPT_STRING_KEY: "I believe the meaning of life is",
+            },
+        ],
+    })
+    assert res.status_code == 200
+    assert type(res.body) == list
+    assert len(res.body) == 2
+    assert res.body[0]["content"] == res.body[1]["content"]
+
     # mixed string and tokens in one sequence
     res = server.make_request("POST", "/completion", data={
         "prompt": [1, 2, 3, 4, 5, 6, prompt_str, 7, 8, 9, 10, prompt_str],
diff --git a/tools/server/tests/unit/test_ctx_shift.py b/tools/server/tests/unit/test_ctx_shift.py
index 2431ac70882d7..8f51bc301a74c 100644
--- a/tools/server/tests/unit/test_ctx_shift.py
+++ b/tools/server/tests/unit/test_ctx_shift.py
@@ -11,7 +11,7 @@
 Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
 """.strip()
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinyllama2()
@@ -25,6 +25,7 @@ def test_ctx_shift_enabled():
     # the prompt is truncated to keep the last 109 tokens
     # 64 tokens are generated thanks to shifting the context when it gets full
     global server
+    server.enable_ctx_shift = True
     server.start()
     res = server.make_request("POST", "/completion", data={
         "n_predict": 64,
@@ -42,7 +43,6 @@ def test_ctx_shift_enabled():
 ])
 def test_ctx_shift_disabled_short_prompt(n_predict: int, n_token_output: int, truncated: bool):
     global server
-    server.disable_ctx_shift = True
     server.n_predict = -1
     server.start()
     res = server.make_request("POST", "/completion", data={
@@ -56,7 +56,6 @@ def test_ctx_shift_disabled_short_prompt(n_predict: int, n_token_output: int, tr
 
 def test_ctx_shift_disabled_long_prompt():
     global server
-    server.disable_ctx_shift = True
     server.start()
     res = server.make_request("POST", "/completion", data={
         "n_predict": 64,
@@ -68,7 +67,6 @@ def test_ctx_shift_disabled_long_prompt():
 
 def test_ctx_shift_disabled_stream():
     global server
-    server.disable_ctx_shift = True
     server.start()
     res = server.make_stream_request("POST", "/v1/completions", data={
         "n_predict": 256,
diff --git a/tools/server/tests/unit/test_embedding.py b/tools/server/tests/unit/test_embedding.py
index 0feb452ccfcd4..50601b839650f 100644
--- a/tools/server/tests/unit/test_embedding.py
+++ b/tools/server/tests/unit/test_embedding.py
@@ -8,7 +8,7 @@
 
 EPSILON = 1e-3
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.bert_bge_small()
diff --git a/tools/server/tests/unit/test_infill.py b/tools/server/tests/unit/test_infill.py
index 10554db0f623e..73dacdae812b8 100644
--- a/tools/server/tests/unit/test_infill.py
+++ b/tools/server/tests/unit/test_infill.py
@@ -3,7 +3,7 @@
 
 server = ServerPreset.tinyllama_infill()
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinyllama_infill()
diff --git a/tools/server/tests/unit/test_lora.py b/tools/server/tests/unit/test_lora.py
index c1aa8be70e2f7..00b2f245f60fc 100644
--- a/tools/server/tests/unit/test_lora.py
+++ b/tools/server/tests/unit/test_lora.py
@@ -5,7 +5,7 @@
 
 LORA_FILE_URL = "https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf"
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.stories15m_moe()
diff --git a/tools/server/tests/unit/test_rerank.py b/tools/server/tests/unit/test_rerank.py
index f4f570ad5ef78..0b63c7821eb98 100644
--- a/tools/server/tests/unit/test_rerank.py
+++ b/tools/server/tests/unit/test_rerank.py
@@ -4,7 +4,7 @@
 server = ServerPreset.jina_reranker_tiny()
 
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.jina_reranker_tiny()
diff --git a/tools/server/tests/unit/test_security.py b/tools/server/tests/unit/test_security.py
index 620b25376bd81..0e11580553aa6 100644
--- a/tools/server/tests/unit/test_security.py
+++ b/tools/server/tests/unit/test_security.py
@@ -6,7 +6,7 @@
 
 TEST_API_KEY = "sk-this-is-the-secret-key"
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinyllama2()
diff --git a/tools/server/tests/unit/test_slot_save.py b/tools/server/tests/unit/test_slot_save.py
index 38704f5ece35a..1b428cc2a840c 100644
--- a/tools/server/tests/unit/test_slot_save.py
+++ b/tools/server/tests/unit/test_slot_save.py
@@ -3,7 +3,7 @@
 
 server = ServerPreset.tinyllama2()
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinyllama2()
diff --git a/tools/server/tests/unit/test_speculative.py b/tools/server/tests/unit/test_speculative.py
index 54db38cf3bd80..38ca4325ba675 100644
--- a/tools/server/tests/unit/test_speculative.py
+++ b/tools/server/tests/unit/test_speculative.py
@@ -16,7 +16,7 @@ def create_server():
     server.draft_max = 8
 
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def fixture_create_server():
     return create_server()
 
@@ -91,6 +91,7 @@ def test_slot_ctx_not_exceeded():
 def test_with_ctx_shift():
     global server
     server.n_ctx = 64
+    server.enable_ctx_shift = True
     server.start()
     res = server.make_request("POST", "/completion", data={
         "prompt": "Hello " * 56,
diff --git a/tools/server/tests/unit/test_tokenize.py b/tools/server/tests/unit/test_tokenize.py
index 382457c9d602f..424cac5f3d394 100644
--- a/tools/server/tests/unit/test_tokenize.py
+++ b/tools/server/tests/unit/test_tokenize.py
@@ -4,7 +4,7 @@
 server = ServerPreset.tinyllama2()
 
 
-@pytest.fixture(scope="module", autouse=True)
+@pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinyllama2()
diff --git a/tools/server/tests/unit/test_tool_call.py b/tools/server/tests/unit/test_tool_call.py
index 20f048c6f6aa5..a3c3ccdf586ab 100755
--- a/tools/server/tests/unit/test_tool_call.py
+++ b/tools/server/tests/unit/test_tool_call.py
@@ -22,6 +22,8 @@ def create_server():
     server.model_alias = "tinyllama-2-tool-call"
     server.server_port = 8081
     server.n_slots = 1
+    server.n_ctx = 8192
+    server.n_batch = 2048
 
 class CompletionMode(Enum):
     NORMAL = "normal"
diff --git a/tools/server/tests/unit/test_vision_api.py b/tools/server/tests/unit/test_vision_api.py
index fc63caa134293..36d14b3885175 100644
--- a/tools/server/tests/unit/test_vision_api.py
+++ b/tools/server/tests/unit/test_vision_api.py
@@ -10,21 +10,48 @@
 
 response = requests.get(IMG_URL_0)
 response.raise_for_status() # Raise an exception for bad status codes
-IMG_BASE64_0 = "data:image/png;base64," + base64.b64encode(response.content).decode("utf-8")
+IMG_BASE64_URI_0 = "data:image/png;base64," + base64.b64encode(response.content).decode("utf-8")
+IMG_BASE64_0 = base64.b64encode(response.content).decode("utf-8")
 
+response = requests.get(IMG_URL_1)
+response.raise_for_status() # Raise an exception for bad status codes
+IMG_BASE64_URI_1 = "data:image/png;base64," + base64.b64encode(response.content).decode("utf-8")
+IMG_BASE64_1 = base64.b64encode(response.content).decode("utf-8")
+
+JSON_MULTIMODAL_KEY = "multimodal_data"
+JSON_PROMPT_STRING_KEY = "prompt_string"
 
 @pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinygemma3()
 
+def test_models_supports_multimodal_capability():
+    global server
+    server.start() # vision model may take longer to load due to download size
+    res = server.make_request("GET", "/models", data={})
+    assert res.status_code == 200
+    model_info = res.body["models"][0]
+    print(model_info)
+    assert "completion" in model_info["capabilities"]
+    assert "multimodal" in model_info["capabilities"]
+
+def test_v1_models_supports_multimodal_capability():
+    global server
+    server.start() # vision model may take longer to load due to download size
+    res = server.make_request("GET", "/v1/models", data={})
+    assert res.status_code == 200
+    model_info = res.body["models"][0]
+    print(model_info)
+    assert "completion" in model_info["capabilities"]
+    assert "multimodal" in model_info["capabilities"]
 
 @pytest.mark.parametrize(
     "prompt, image_url, success, re_content",
     [
         # test model is trained on CIFAR-10, but it's quite dumb due to small size
         ("What is this:\n", IMG_URL_0,                True, "(cat)+"),
-        ("What is this:\n", "IMG_BASE64_0",           True, "(cat)+"), # exceptional, so that we don't cog up the log
+        ("What is this:\n", "IMG_BASE64_URI_0",       True, "(cat)+"), # exceptional, so that we don't cog up the log
         ("What is this:\n", IMG_URL_1,                True, "(frog)+"),
         ("Test test\n",     IMG_URL_1,                True, "(frog)+"), # test invalidate cache
         ("What is this:\n", "malformed",              False, None),
@@ -36,8 +63,8 @@ def create_server():
 def test_vision_chat_completion(prompt, image_url, success, re_content):
     global server
     server.start(timeout_seconds=60) # vision model may take longer to load due to download size
-    if image_url == "IMG_BASE64_0":
-        image_url = IMG_BASE64_0
+    if image_url == "IMG_BASE64_URI_0":
+        image_url = IMG_BASE64_URI_0
     res = server.make_request("POST", "/chat/completions", data={
         "temperature": 0.0,
         "top_k": 1,
@@ -58,3 +85,61 @@ def test_vision_chat_completion(prompt, image_url, success, re_content):
     else:
         assert res.status_code != 200
 
+
+@pytest.mark.parametrize(
+    "prompt, image_data, success, re_content",
+    [
+        # test model is trained on CIFAR-10, but it's quite dumb due to small size
+        ("What is this: <__media__>\n", IMG_BASE64_0,           True, "(cat)+"),
+        ("What is this: <__media__>\n", IMG_BASE64_1,           True, "(frog)+"),
+        ("What is this: <__media__>\n", "malformed",            False, None), # non-image data
+        ("What is this:\n",             "",                     False, None), # empty string
+    ]
+)
+def test_vision_completion(prompt, image_data, success, re_content):
+    global server
+    server.start() # vision model may take longer to load due to download size
+    res = server.make_request("POST", "/completions", data={
+        "temperature": 0.0,
+        "top_k": 1,
+        "prompt": { JSON_PROMPT_STRING_KEY: prompt, JSON_MULTIMODAL_KEY: [ image_data ] },
+    })
+    if success:
+        assert res.status_code == 200
+        content = res.body["content"]
+        assert match_regex(re_content, content)
+    else:
+        assert res.status_code != 200
+
+
+@pytest.mark.parametrize(
+    "prompt, image_data, success",
+    [
+        # test model is trained on CIFAR-10, but it's quite dumb due to small size
+        ("What is this: <__media__>\n", IMG_BASE64_0,           True), # exceptional, so that we don't cog up the log
+        ("What is this: <__media__>\n", IMG_BASE64_1,           True),
+        ("What is this: <__media__>\n", "malformed",            False), # non-image data
+        ("What is this:\n",             "base64",               False), # non-image data
+    ]
+)
+def test_vision_embeddings(prompt, image_data, success):
+    global server
+    server.server_embeddings=True
+    server.n_batch=512
+    server.start() # vision model may take longer to load due to download size
+    res = server.make_request("POST", "/embeddings", data={
+        "content": [
+            { JSON_PROMPT_STRING_KEY: prompt, JSON_MULTIMODAL_KEY: [ image_data ] },
+            { JSON_PROMPT_STRING_KEY: prompt, JSON_MULTIMODAL_KEY: [ image_data ] },
+            { JSON_PROMPT_STRING_KEY: prompt, },
+        ],
+    })
+    if success:
+        assert res.status_code == 200
+        content = res.body
+        # Ensure embeddings are stable when multimodal.
+        assert content[0]['embedding'] == content[1]['embedding']
+        # Ensure embeddings without multimodal but same prompt do not match multimodal embeddings.
+        assert content[0]['embedding'] != content[2]['embedding']
+    else:
+        assert res.status_code != 200
diff --git a/tools/server/tests/utils.py b/tools/server/tests/utils.py
index bc547ca03bf1b..5f42bcae61e1a 100644
--- a/tools/server/tests/utils.py
+++ b/tools/server/tests/utils.py
@@ -79,7 +79,7 @@ class ServerProcess:
     draft: int | None = None
     api_key: str | None = None
     lora_files: List[str] | None = None
-    disable_ctx_shift: int | None = False
+    enable_ctx_shift: int | None = False
     draft_min: int | None = None
     draft_max: int | None = None
     no_webui: bool | None = None
@@ -178,8 +178,8 @@ def start(self, timeout_seconds: int | None = DEFAULT_HTTP_TIMEOUT) -> None:
         if self.lora_files:
             for lora_file in self.lora_files:
                 server_args.extend(["--lora", lora_file])
-        if self.disable_ctx_shift:
-            server_args.extend(["--no-context-shift"])
+        if self.enable_ctx_shift:
+            server_args.append("--context-shift")
         if self.api_key:
             server_args.extend(["--api-key", self.api_key])
         if self.draft_max:
@@ -318,46 +318,53 @@ def make_any_request(
             arguments_parts = 0
 
             for chunk in self.make_stream_request(method, path, data, headers):
-                assert len(chunk['choices']) == 1, f'Expected 1 choice, got {len(chunk["choices"])}'
-                choice = chunk['choices'][0]
-                if choice['delta'].get('content') is not None:
-                    assert len(choice['delta']['content']) > 0, f'Expected non empty content delta!'
-                    content.append(choice['delta']['content'])
-                    content_parts += 1
-                if choice['delta'].get('reasoning_content') is not None:
-                    assert len(choice['delta']['reasoning_content']) > 0, f'Expected non empty reasoning_content delta!'
-                    reasoning_content.append(choice['delta']['reasoning_content'])
-                    reasoning_content_parts += 1
-                if choice['delta'].get('finish_reason') is not None:
-                    finish_reason = choice['delta']['finish_reason']
-                for tc in choice['delta'].get('tool_calls', []):
-                    if 'function' not in tc:
-                        raise ValueError(f"Expected function type, got {tc['type']}")
-                    if tc['index'] >= len(tool_calls):
-                        assert 'id' in tc
-                        assert tc.get('type') == 'function'
-                        assert 'function' in tc and 'name' in tc['function'] and len(tc['function']['name']) > 0, \
-                            f"Expected function call with name, got {tc.get('function')}"
-                        tool_calls.append(dict(
-                            id="",
-                            type="function",
-                            function=dict(
-                                name="",
-                                arguments="",
-                            )
-                        ))
-                    tool_call = tool_calls[tc['index']]
-                    if tc.get('id') is not None:
-                        tool_call['id'] = tc['id']
-                    fct = tc['function']
-                    assert 'id' not in fct, f"Function call should not have id: {fct}"
-                    if fct.get('name') is not None:
-                        tool_call['function']['name'] = tool_call['function'].get('name', '') + fct['name']
-                    if fct.get('arguments') is not None:
-                        tool_call['function']['arguments'] += fct['arguments']
-                        arguments_parts += 1
-                    tool_call_parts += 1
-
+                if chunk['choices']:
+                    assert len(chunk['choices']) == 1, f'Expected 1 choice, got {len(chunk["choices"])}'
+                    choice = chunk['choices'][0]
+                    if choice['delta'].get('content') is not None:
+                        assert len(choice['delta']['content']) > 0, f'Expected non empty content delta!'
+                        content.append(choice['delta']['content'])
+                        content_parts += 1
+                    if choice['delta'].get('reasoning_content') is not None:
+                        assert len(choice['delta']['reasoning_content']) > 0, f'Expected non empty reasoning_content delta!'
+                        reasoning_content.append(choice['delta']['reasoning_content'])
+                        reasoning_content_parts += 1
+                    if choice['delta'].get('finish_reason') is not None:
+                        finish_reason = choice['delta']['finish_reason']
+                    for tc in choice['delta'].get('tool_calls', []):
+                        if 'function' not in tc:
+                            raise ValueError(f"Expected function type, got {tc['type']}")
+                        if tc['index'] >= len(tool_calls):
+                            assert 'id' in tc
+                            assert tc.get('type') == 'function'
+                            assert 'function' in tc and 'name' in tc['function'] and len(tc['function']['name']) > 0, \
+                                f"Expected function call with name, got {tc.get('function')}"
+                            tool_calls.append(dict(
+                                id="",
+                                type="function",
+                                function=dict(
+                                    name="",
+                                    arguments="",
+                                )
+                            ))
+                        tool_call = tool_calls[tc['index']]
+                        if tc.get('id') is not None:
+                            tool_call['id'] = tc['id']
+                        fct = tc['function']
+                        assert 'id' not in fct, f"Function call should not have id: {fct}"
+                        if fct.get('name') is not None:
+                            tool_call['function']['name'] = tool_call['function'].get('name', '') + fct['name']
+                        if fct.get('arguments') is not None:
+                            tool_call['function']['arguments'] += fct['arguments']
+                            arguments_parts += 1
+                        tool_call_parts += 1
+                else:
+                    # When `include_usage` is True (the default), we expect the last chunk of the stream
+                    # immediately preceding the `data: [DONE]` message to contain a `choices` field with an empty array
+                    # and a `usage` field containing the usage statistics (n.b., llama-server also returns `timings` in
+                    # the last chunk)
+                    assert 'usage' in chunk, f"Expected finish_reason in chunk: {chunk}"
+                    assert 'timings' in chunk, f"Expected finish_reason in chunk: {chunk}"
             print(f'Streamed response had {content_parts} content parts, {reasoning_content_parts} reasoning_content parts, {tool_call_parts} tool call parts incl. {arguments_parts} arguments parts')
             result = dict(
                 choices=[
diff --git a/tools/server/utils.hpp b/tools/server/utils.hpp
index f3dfc8225da4d..036060bb3e9c7 100644
--- a/tools/server/utils.hpp
+++ b/tools/server/utils.hpp
@@ -123,6 +123,19 @@ static bool json_is_array_of_mixed_numbers_strings(const json & data) {
     return false;
 }
 
+// does array have any individual integers/tokens?
+static bool json_is_array_and_contains_numbers(const json & data) {
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            if (e.is_number_integer()) {
+                return true;
+            }
+        }
+        return false;
+    }
+    return false;
+}
+
 // get value by path(key1 / key2)
 static json json_get_nested_values(const std::vector<std::string> & paths, const json & js) {
     json result = json::object();
@@ -186,48 +199,6 @@ static llama_tokens tokenize_mixed(const llama_vocab * vocab, const json & json_
     return prompt_tokens;
 }
 
-/**
- * break the input "prompt" object into multiple prompt if needed, then tokenize them
- * this supports these cases:
- * - "prompt": "string"
- * - "prompt": [12, 34, 56]
- * - "prompt": [12, 34, "string", 56, 78]
- * and multiple prompts (multi-tasks):
- * - "prompt": ["string1", "string2"]
- * - "prompt": ["string1", [12, 34, 56]]
- * - "prompt": [[12, 34, 56], [78, 90, 12]]
- * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
- */
-static std::vector<llama_tokens> tokenize_input_prompts(const llama_vocab * vocab, const json & json_prompt, bool add_special, bool parse_special) {
-    std::vector<llama_tokens> result;
-    if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
-        // string or mixed
-        result.push_back(tokenize_mixed(vocab, json_prompt, add_special, parse_special));
-    } else if (json_is_array_of_numbers(json_prompt)) {
-        // array of tokens
-        result.push_back(json_prompt.get<llama_tokens>());
-    } else if (json_prompt.is_array()) {
-        // array of prompts
-        result.reserve(json_prompt.size());
-        for (const auto & p : json_prompt) {
-            if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
-                result.push_back(tokenize_mixed(vocab, p, add_special, parse_special));
-            } else if (json_is_array_of_numbers(p)) {
-                // array of tokens
-                result.push_back(p.get<llama_tokens>());
-            } else {
-                throw std::runtime_error("element of \"prompt\" must be a string, an list of tokens, or a list of mixed strings & tokens");
-            }
-        }
-    } else {
-        throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts");
-    }
-    if (result.empty()) {
-        throw std::runtime_error("\"prompt\" must not be empty");
-    }
-    return result;
-}
-
 // return the last index of character that can form a valid string
 // if the last character is potentially cut in half, return the index before the cut
 // if validate_utf8(text) == text.size(), then the whole text is valid utf8
@@ -262,35 +233,6 @@ static size_t validate_utf8(const std::string& text) {
 // template utils
 //
 
-// format rerank task: [BOS]query[EOS][SEP]doc[EOS]
-static llama_tokens format_rerank(const struct llama_vocab * vocab, const llama_tokens & query, const llama_tokens & doc) {
-    llama_tokens result;
-
-    // Get EOS token - use SEP token as fallback if EOS is not available
-    llama_token eos_token = llama_vocab_eos(vocab);
-    if (eos_token == LLAMA_TOKEN_NULL) {
-        eos_token = llama_vocab_sep(vocab);
-    }
-
-    result.reserve(doc.size() + query.size() + 4);
-    if (llama_vocab_get_add_bos(vocab)) {
-        result.push_back(llama_vocab_bos(vocab));
-    }
-    result.insert(result.end(), query.begin(), query.end());
-    if (llama_vocab_get_add_eos(vocab)) {
-        result.push_back(eos_token);
-    }
-    if (llama_vocab_get_add_sep(vocab)) {
-        result.push_back(llama_vocab_sep(vocab));
-    }
-    result.insert(result.end(), doc.begin(), doc.end());
-    if (llama_vocab_get_add_eos(vocab)) {
-        result.push_back(eos_token);
-    }
-
-    return result;
-}
-
 // format infill task
 static llama_tokens format_infill(
         const llama_vocab * vocab,
@@ -1186,6 +1128,24 @@ struct server_tokens {
         }
     }
 
+    // appends server tokens, updates the media map. copies media chunks.
+    void push_back(server_tokens & tokens) {
+        size_t start_pos = size();
+        for (size_t i = 0; i < tokens.size(); i++) {
+            push_back(tokens[i]);
+        }
+        if (tokens.has_mtmd) {
+            // Assert if we are copying MTMD chunks to a server_tokens that does not have mtmd.
+            // We could also just check, but this will prevent silently dropping MTMD data.
+            GGML_ASSERT(has_mtmd);
+            for (auto it = tokens.map_pos_to_media.begin(); it != tokens.map_pos_to_media.end(); ) {
+                auto chunk = tokens.map_pos_to_media[it->first].get();
+                mtmd::input_chunk_ptr new_chunk(mtmd_input_chunk_copy(chunk));
+                map_pos_to_media[start_pos+it->first] = std::move(new_chunk);
+            }
+        }
+    }
+
     // for compatibility with context shift and prompt truncation
     void insert(const llama_tokens & inp_tokens) {
         GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled
@@ -1356,3 +1316,137 @@ static std::string fnv_hash(const uint8_t * data, size_t len) {
     }
     return std::to_string(hash);
 }
+
+
+// format rerank task: [BOS]query[EOS][SEP]doc[EOS].
+static server_tokens format_rerank(const struct llama_vocab * vocab, server_tokens & query, server_tokens & doc) {
+    server_tokens result = {};
+
+    // Get EOS token - use SEP token as fallback if EOS is not available
+    llama_token eos_token = llama_vocab_eos(vocab);
+    if (eos_token == LLAMA_TOKEN_NULL) {
+        eos_token = llama_vocab_sep(vocab);
+    }
+    if (llama_vocab_get_add_bos(vocab)) {
+        result.push_back(llama_vocab_bos(vocab));
+    }
+    result.push_back(query);
+    if (llama_vocab_get_add_eos(vocab)) {
+        result.push_back(eos_token);
+    }
+    if (llama_vocab_get_add_sep(vocab)) {
+        result.push_back(llama_vocab_sep(vocab));
+    }
+    result.push_back(doc);
+    if (llama_vocab_get_add_eos(vocab)) {
+        result.push_back(eos_token);
+    }
+    return result;
+}
+
+
+static server_tokens process_mtmd_prompt(mtmd_context * mctx, std::string prompt, std::vector<raw_buffer> files) {
+    mtmd::bitmaps bitmaps;
+    for (auto & file : files) {
+        mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(mctx, file.data(), file.size()));
+        if (!bmp.ptr) {
+            throw std::runtime_error("Failed to load image or audio file");
+        }
+        // calculate bitmap hash (for KV caching)
+        std::string hash = fnv_hash(bmp.data(), bmp.n_bytes());
+        bmp.set_id(hash.c_str());
+        bitmaps.entries.push_back(std::move(bmp));
+    }
+    // process prompt
+    std::vector<server_tokens> inputs;
+    // multimodal
+    mtmd_input_text inp_txt = {
+        prompt.c_str(),
+        /* add_special */   true,
+        /* parse_special */ true,
+    };
+    mtmd::input_chunks chunks(mtmd_input_chunks_init());
+    auto bitmaps_c_ptr = bitmaps.c_ptr();
+    int32_t tokenized = mtmd_tokenize(mctx,
+                                      chunks.ptr.get(),
+                                      &inp_txt,
+                                      bitmaps_c_ptr.data(),
+                                      bitmaps_c_ptr.size());
+    if (tokenized != 0) {
+        throw std::runtime_error("Failed to tokenize prompt");
+    }
+    auto result = server_tokens(chunks, true);
+    return result;
+}
+
+/**
+ * break the input "prompt" object into multiple prompt if needed, then tokenize them
+ * use tokenize_input_prompts() if the input could be an array.
+ * this supports these cases:
+ * - "prompt": "string"
+ * - "prompt": [12, 34, 56]
+ * - "prompt": [12, 34, "string", 56, 78]
+ * - "prompt": { "prompt_string": "string", "multimodal_data": [ "base64" ] }
+ */
+static server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special) {
+    constexpr char JSON_STRING_PROMPT_KEY[] = "prompt_string";
+    constexpr char JSON_MTMD_DATA_KEY[] = "multimodal_data";
+    const bool has_mtmd = mctx != nullptr;
+    if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
+        // string or mixed
+        llama_tokens tmp = tokenize_mixed(vocab, json_prompt, add_special, parse_special);
+        return server_tokens(tmp, false);
+    } else if (json_is_array_of_numbers(json_prompt)) {
+        // array of tokens
+        llama_tokens tmp = json_prompt.get<llama_tokens>();
+        return server_tokens(tmp, false);
+    } else if (json_prompt.contains(JSON_STRING_PROMPT_KEY)) {
+        // JSON object with prompt key.
+        if (json_prompt.contains(JSON_MTMD_DATA_KEY)) {
+            if (!has_mtmd)
+                throw std::runtime_error("Multimodal data provided, but model does not support multimodal requests.");
+
+            // JSON object with prompt and multimodal key.
+            std::vector<raw_buffer> files;
+            for (const auto & entry : json_prompt.at(JSON_MTMD_DATA_KEY)) {
+                files.push_back(base64_decode(entry));
+            }
+            return process_mtmd_prompt(mctx, json_prompt.at(JSON_STRING_PROMPT_KEY), files);
+        } else {
+            // Not multimodal, but contains a subobject.
+            llama_tokens tmp = tokenize_mixed(vocab, json_prompt.at(JSON_STRING_PROMPT_KEY), add_special, parse_special);
+            return server_tokens(tmp, false);
+        }
+   } else {
+       throw std::runtime_error("\"prompt\" elements must be a string, a list of tokens, a JSON object containing a prompt string, or a list of mixed strings & tokens.");
+   }
+}
+
+/**
+ * break the input "prompt" object into multiple prompt if needed, then tokenize them
+ * this supports these cases:
+ * - "prompt": "string"
+ * - "prompt": [12, 34, 56]
+ * - "prompt": [12, 34, "string", 56, 78]
+ * - "prompt": { "prompt_string": "string", "multimodal_data": [ "base64" ] }
+ * and multiple prompts (multi-tasks):
+ * - "prompt": ["string1", "string2"]
+ * - "prompt": ["string1", [12, 34, 56]]
+ * - "prompt": [[12, 34, 56], [78, 90, 12]]
+ * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56], { "prompt_string": "string", "multimodal_data": [ "base64" ]}]
+ */
+static std::vector<server_tokens> tokenize_input_prompts(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special) {
+    std::vector<server_tokens> result;
+    if (json_prompt.is_array() && !json_is_array_and_contains_numbers(json_prompt)) {
+        result.reserve(json_prompt.size());
+        for (const auto & p : json_prompt) {
+            result.push_back(tokenize_input_subprompt(vocab, mctx, p,add_special, parse_special));
+        }
+    } else {
+        result.push_back(tokenize_input_subprompt(vocab, mctx, json_prompt, add_special, parse_special));
+    }
+    if (result.empty()) {
+        throw std::runtime_error("\"prompt\" must not be empty");
+    }
+    return result;
+}
diff --git a/tools/server/webui/src/components/ChatMessage.tsx b/tools/server/webui/src/components/ChatMessage.tsx
index 1c579f2659050..a1a7d152f9e78 100644
--- a/tools/server/webui/src/components/ChatMessage.tsx
+++ b/tools/server/webui/src/components/ChatMessage.tsx
@@ -62,8 +62,7 @@ export default function ChatMessage({
       return { content: msg.content };
     }
     const REGEX_THINK_OPEN = /<think>|<\|channel\|>analysis<\|message\|>/;
-    const REGEX_THINK_CLOSE =
-      /<\/think>|<\|start\|>assistant<\|channel\|>final<\|message\|>/;
+    const REGEX_THINK_CLOSE = /<\/think>|<\|end\|>/;
     let actualContent = '';
     let thought = '';
     let isThinking = false;
diff --git a/tools/server/webui/src/utils/app.context.tsx b/tools/server/webui/src/utils/app.context.tsx
index 96cffd95aba7c..1860ecc721821 100644
--- a/tools/server/webui/src/utils/app.context.tsx
+++ b/tools/server/webui/src/utils/app.context.tsx
@@ -209,6 +209,7 @@ export const AppContextProvider = ({
         messages,
         stream: true,
         cache_prompt: true,
+        reasoning_format: 'none',
         samplers: config.samplers,
         temperature: config.temperature,
         dynatemp_range: config.dynatemp_range,
@@ -254,7 +255,7 @@ export const AppContextProvider = ({
         if (chunk.error) {
           throw new Error(chunk.error?.message || 'Unknown error');
         }
-        const addedContent = chunk.choices[0].delta.content;
+        const addedContent = chunk.choices[0]?.delta.content;
         const lastContent = pendingMsg.content || '';
         if (addedContent) {
           pendingMsg = {
diff --git a/tools/server/webui/src/utils/misc.ts b/tools/server/webui/src/utils/misc.ts
index d60a68cd2431b..15182e4cab436 100644
--- a/tools/server/webui/src/utils/misc.ts
+++ b/tools/server/webui/src/utils/misc.ts
@@ -130,7 +130,10 @@ export function filterThoughtFromMsgs(messages: APIMessage[]) {
       role: msg.role,
       content:
         msg.role === 'assistant'
-          ? contentStr.split('</think>').at(-1)!.trim()
+          ? contentStr
+              .split(/<\/think>|<\|end\|>/)
+              .at(-1)!
+              .trim()
           : contentStr,
     } as APIMessage;
   });
diff --git a/tools/tts/tts.cpp b/tools/tts/tts.cpp
index a71e9bf5b589e..18f01a9946350 100644
--- a/tools/tts/tts.cpp
+++ b/tools/tts/tts.cpp
@@ -581,7 +581,6 @@ int main(int argc, char ** argv) {
 
     params.model = params.vocoder.model;
     params.embedding = true;
-    params.ctx_shift = false; // silence warning
     params.n_ubatch = params.n_batch;
 
     common_init_result llama_init_cts = common_init_from_params(params);