Update on "[ET-VK][qlinear] Faster weight only quantized linear gemv kernel"

## Changes

* Introduce a new compute shader for int4 linear's gemv cases that performs much better than the existing shader. This shader is inspired from MNN's gemv_1x1_conv_buf.cl shader.

With this compute kernel, transformer models' text generation can execute much faster than before.

On Samsung Galaxy S24 for Llama 3.2 1B, generating 128 tokens:

Before: ~25 tok/s
After: ~49 tok/s

## Why this new shader is faster

The biggest reason is due to vectorized loading of the uint4 weight buffer. This new shader loads the weight buffer as a buffer/image of `uvec4`, whereas the old shader loads the weight buffer as a buffer/image of `u8vec4`. Using the Adreno Offline Compiler, I found that in the former, only one load instruction was used to load from the weight tensor, whereas in the latter 16 load instructions were used to load from the weight tensor. It appears that the data loading was not being vectorized at the assembly level. This is potentially behaviour that can be approved in the SPIR-V shader compiler.

An additional factor is better weight packing layout. The new prepacking routine results in better memory coalescing between threads in a work group.

The final major factor is the use of tree based reduction to co-operatively reduce partial results into the final output. Previously, a single thread was responsible for the final reduction.

## Future Work

* Introduce faster shader for int4 linear gemm cases
* Update QCSNW to also use these updated shaders

Differential Revision: [D78275584](https://our.internmc.facebook.com/intern/diff/D78275584/)

[ghstack-poisoned]

Author: SS-JIA

.ci/scripts/unittest-buck2.sh

@@ -10,7 +10,7 @@ set -eux
 # TODO: can't query cadence & vulkan backends
 # TODO: can't query //kernels/prim_ops because of non-buckified stuff in OSS.
 buck2 query "//backends/apple/... + //backends/example/... + \
-//backends/mediatek/... + //backends/test/... + //backends/transforms/... + \
+//backends/mediatek/... + //backends/transforms/... + \
 //backends/xnnpack/... + //configurations/... + //kernels/aten/... + \
 //kernels/optimized/... + //kernels/portable/... + //kernels/quantized/... + \
 //kernels/test/... + //runtime/... + //schema/... + //test/... + //util/..."

.github/workflows/android-release-artifacts.yml

@@ -47,15 +47,13 @@ jobs:
     name: build-aar
     needs: check-if-aar-exists
     if: ${{ !github.event.pull_request.head.repo.fork }}
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@release/2.7
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     secrets: inherit
     permissions:
       id-token: write
       contents: read
     with:
       secrets-env: EXECUTORCH_MAVEN_SIGNING_KEYID EXECUTORCH_MAVEN_SIGNING_PASSWORD EXECUTORCH_MAVEN_CENTRAL_PASSWORD EXECUTORCH_MAVEN_CENTRAL_USERNAME EXECUTORCH_MAVEN_SIGNING_GPG_KEY_CONTENTS
-      # As this job has access to Maven credential, run this on a fresh ephemeral runner
-      runner: ephemeral.linux.2xlarge
       docker-image: executorch-ubuntu-22.04-clang12-android
       submodules: 'recursive'
       ref: ${{ github.sha }}

backends/arm/scripts/mlsdk_utils.sh

@@ -0,0 +1,121 @@
+#!/usr/bin/env bash
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -euo pipefail
+
+# TODO
+mlsdk_manifest_url=""
+
+script_dir=$(cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd)
+
+source ${script_dir}/utils.sh
+
+usage() { echo "Usage: $0 [-u <mlsdk-manifest-url>]" 1>&2; exit 1; }
+
+while getopts ":u:" opt; do
+    case "${opt}" in
+        u)
+            mlsdk_manifest_url=${OPTARG}
+            ;;
+        *)
+            usage
+            ;;
+    esac
+done
+
+function download_ai_mlsdk_manifest() {
+    local _dada_dir="$1"
+
+    if [[ -z "${_dada_dir}" ]]; then
+        echo "Error: _dada_dir parameter missing?"
+        return 1
+    fi
+
+    if [[ -z "${mlsdk_manifest_url}" ]]; then
+        echo "Error: mlsdk_manifest_url parameter missing?"
+        return 1
+    fi
+
+    if [[ ! -d "${_dada_dir}" ]]; then
+        mkdir -p "$_dada_dir"
+        pushd "$_dada_dir" || exit 1
+
+        curl https://storage.googleapis.com/git-repo-downloads/repo > repo
+        chmod u+x repo
+        ./repo init  --no-repo-verify --depth=1  --manifest-url  ${mlsdk_manifest_url} -g model-converter,emulation-layer,vgf-library
+        ./repo sync
+
+        popd
+    fi
+}
+
+function setup_model_converter() {
+    local work_dir="$1"
+    local manifest_dir="$2"
+    local enable_vgf_lib="$3"
+    local enable_emulation_layer="$4"
+
+    if [[ -z "$work_dir" ]]; then
+        echo "Error: work_dir parameter is required."
+        return 1
+    fi
+
+    if [[ -z "$manifest_dir" ]]; then
+        echo "Error: manifest_dir parameter is required."
+        return 1
+    fi
+
+    mkdir -p "$work_dir"
+    pushd "$work_dir" || exit 1
+
+    download_ai_mlsdk_manifest ${manifest_dir}
+
+    pushd "$manifest_dir"
+
+    # model-converter
+    # TODO: Remove macOS patch after mlsdk fully supports macOS
+    if [[ "$(uname)" == "Darwin" ]]; then
+    sed -i '' '/^ *print(f"Unsupported host platform/ i\
+            if system == "Darwin":\
+                # Use default Apple toolchain (Clang) on macOS\
+                return True\
+\
+' sw/model-converter/scripts/build.py
+    fi
+    python sw/model-converter/scripts/build.py -j$(nproc)
+
+    # libvgf
+    if [[ "${enable_vgf_lib}" -eq 1 ]]; then
+    # TODO: Remove macOS patch after mlsdk fully supports macOS
+        if [[ "$(uname)" == "Darwin" ]]; then
+    sed -i '' '/^ *print(f"ERROR: Unsupported host platform/ i\
+            if system == "Darwin":\
+                # Use default Apple toolchain (Clang) on macOS\
+                return True\
+\
+' sw/vgf-lib/scripts/build.py
+        fi
+        python sw/vgf-lib/scripts/build.py -j$(nproc)
+    fi
+
+    # emu layer
+    if [[ "${enable_emulation_layer}" -eq 1 ]]; then
+        pushd sw/emulation-layer
+        cmake -B build                                               \
+            -DGLSLANG_PATH=../../dependencies/glslang                \
+            -DSPIRV_CROSS_PATH=../../dependencies/SPIRV-Cross        \
+            -DSPIRV_HEADERS_PATH=../../dependencies/SPIRV-Headers    \
+            -DSPIRV_TOOLS_PATH=../../dependencies/SPIRV-Tools        \
+            -DVULKAN_HEADERS_PATH=../../dependencies/Vulkan-Headers
+        cmake --build build
+        popd
+    fi
+
+    popd
+}
+
+#setup_model_converter() $1
+# `"$manifest_dir"'

backends/arm/tosa_quant_utils.py

@@ -146,7 +146,15 @@ def quantize_value(self, x: torch.Tensor | float) -> Tensor:
         x = x.to(torch.float32)
         if self.per_channel:
             q_op = exir_ops.edge.quantized_decomposed.quantize_per_channel.default
-            args = (x, self.scale, self.zp, self.axis, self.qmin, self.qmax, self.dtype)
+            args = (
+                x,
+                torch.tensor(self.scale),
+                torch.tensor(self.zp),
+                self.axis,
+                self.qmin,
+                self.qmax,
+                self.dtype,
+            )
         else:
             q_op = exir_ops.edge.quantized_decomposed.quantize_per_tensor.default
             args = (x, self.scale, self.zp, self.qmin, self.qmax, self.dtype)  # type: ignore[assignment]
@@ -162,8 +170,8 @@ def dequantize_value(self, qx: torch.Tensor) -> torch.Tensor:
             dq_op = exir_ops.edge.quantized_decomposed.dequantize_per_channel.default
             args = (
                 qx,
-                self.scale,
-                self.zp,
+                torch.tensor(self.scale),
+                torch.tensor(self.zp),
                 self.axis,
                 self.qmin,
                 self.qmax,

backends/cadence/aot/TARGETS

@@ -184,6 +184,21 @@ python_library(
     ],
 )
 
+python_library(
+    name = "program_builder",
+    srcs = [
+        "program_builder.py",
+    ],
+    typing = True,
+    deps = [
+        ":graph_builder",
+        "fbcode//caffe2:torch",
+        "fbcode//executorch/exir:lib",
+        "fbcode//executorch/exir:pass_base",
+        "fbcode//executorch/exir/verification:verifier",
+    ],
+)
+
 python_library(
     name = "fuse_ops",
     srcs = [
@@ -508,6 +523,7 @@ python_unittest(
         ":typing_stubs",
         ":ops_registrations",
         ":pass_utils",
+        ":program_builder",
         "//caffe2:torch",
         "//executorch/exir:memory",
         "//executorch/exir/dialects:lib",