Merge branch 'main' into op-floor-div

Author: zingo

.ci/docker/requirements-ci.txt

@@ -1,7 +1,7 @@
 mpmath==1.3.0
 numpy>=2.0.0; python_version >= '3.10'
 PyYAML==6.0.1
-ruamel.yaml==0.17.32
+ruamel.yaml==0.18.15
 sympy==1.12
 timm==0.6.13
 tomli==2.0.1

.ci/scripts/test-cuda-build.sh

@@ -27,9 +27,6 @@ test_executorch_cuda_build() {
     nvcc --version || echo "nvcc not found"
     nvidia-smi || echo "nvidia-smi not found"
 
-    # Set CMAKE_ARGS to enable CUDA build - ExecuTorch will handle PyTorch installation automatically
-    export CMAKE_ARGS="-DEXECUTORCH_BUILD_CUDA=ON"
-
     echo "=== Starting ExecuTorch Installation ==="
     # Install ExecuTorch with CUDA support with timeout and error handling
     timeout 5400 ./install_executorch.sh || {

.ci/scripts/test_llama_lora.sh

@@ -55,7 +55,7 @@ cmake_build_llama_runner
 # Constants.
 RUNTIME_ARGS="--tokenizer_path=${DOWNLOADED_PATH}/tokenizer.model --temperature=0 --seq_len=20 --warmup=1"
 PROMPT="What happens if you eat watermelon seeds?"
-EXPECTED_PREFIX="What happens if you eat watermelon seeds? Watermelon seeds are a good source of vitamin C,"
+EXPECTED_PREFIX="What happens if you eat watermelon seeds? Watermelon seeds are a good source of vitamin C and"
 
 # Export LoRA PTE file.
 MODEL_NAME="llama_3_2_1B_lora"
@@ -94,7 +94,7 @@ else
   exit 1
 fi
 
-# Export LoRA PTE, PTD file.
+# Export LoRA PTE, foundation PTD file.
 MODEL_SEPARATE="${MODEL_NAME}_separate"
 $PYTHON_EXECUTABLE -m extension.llm.export.export_llm \
     base.checkpoint="${DOWNLOADED_PATH}/consolidated.00.pth" \
@@ -114,20 +114,62 @@ $PYTHON_EXECUTABLE -m extension.llm.export.export_llm \
 NOW=$(date +"%H:%M:%S")
 echo "Starting to run llama runner at ${NOW}"
 # shellcheck source=/dev/null
-cmake-out/examples/models/llama/llama_main --model_path=${MODEL_SEPARATE}.pte --data_path=${MODEL_SEPARATE}.ptd --prompt="${PROMPT}" ${RUNTIME_ARGS} > result2.txt
+cmake-out/examples/models/llama/llama_main --model_path=${MODEL_SEPARATE}.pte --data_paths=${MODEL_SEPARATE}.ptd --prompt="${PROMPT}" ${RUNTIME_ARGS} > result2.txt
 NOW=$(date +"%H:%M:%S")
 echo "Finished at ${NOW}"
 
 RESULT2=$(cat result2.txt)
 if [[ "${RESULT2}" == "${EXPECTED_PREFIX}"* ]]; then
   echo "Expected result prefix: ${EXPECTED_PREFIX}"
   echo "Actual result: ${RESULT2}"
+  # Do not clean up files if test passes, as they're re-used in the next test.
   echo "Success"
-  cleanup_files
 else
   echo "Expected result prefix: ${EXPECTED_PREFIX}"
   echo "Actual result: ${RESULT2}"
   echo "Failure; results not the same"
   cleanup_files
   exit 1
 fi
+
+# Export LoRA PTE, LoRA PTD, foundation PTD file.
+MODEL_PROGRAM_ONLY="${MODEL_NAME}_program"
+MODEL_LORA_WEIGHTS="lora_weights"
+MODEL_FOUNDATION_WEIGHTS="foundation_weights"
+$PYTHON_EXECUTABLE -m extension.llm.export.export_llm \
+    base.checkpoint="${DOWNLOADED_PATH}/consolidated.00.pth" \
+    base.params="${DOWNLOADED_PATH}/params.json" \
+    base.adapter_checkpoint="${DOWNLOADED_PATH}/adapter_model.pt" \
+    base.adapter_config="${DOWNLOADED_PATH}/adapter_config.json" \
+    base.tokenizer_path="${DOWNLOADED_PATH}/tokenizer.model" \
+    model.use_kv_cache=true \
+    model.use_sdpa_with_kv_cache=true \
+    model.dtype_override="fp32" \
+    backend.xnnpack.enabled=true \
+    backend.xnnpack.extended_ops=true \
+    export.output_name="${MODEL_PROGRAM_ONLY}.pte" \
+    export.foundation_weights_file="${MODEL_FOUNDATION_WEIGHTS}.ptd" \
+    export.lora_weights_file="${MODEL_LORA_WEIGHTS}.ptd"
+
+# Run llama runner.
+NOW=$(date +"%H:%M:%S")
+echo "Starting to run llama runner at ${NOW}"
+# shellcheck source=/dev/null
+cmake-out/examples/models/llama/llama_main --model_path=${MODEL_PROGRAM_ONLY}.pte --data_paths="${MODEL_FOUNDATION_WEIGHTS}.ptd,${MODEL_LORA_WEIGHTS}.ptd" --prompt="${PROMPT}" ${RUNTIME_ARGS} > result3.txt
+NOW=$(date +"%H:%M:%S")
+echo "Finished at ${NOW}"
+
+RESULT3=$(cat result3.txt)
+if [[ "${RESULT3}" == "${EXPECTED_PREFIX}"* ]]; then
+  echo "Expected result prefix: ${EXPECTED_PREFIX}"
+  echo "Actual result: ${RESULT3}"
+  echo "Success"
+else
+  echo "Expected result prefix: ${EXPECTED_PREFIX}"
+  echo "Actual result: ${RESULT3}"
+  echo "Failure; results not the same"
+  cleanup_files
+  exit 1
+fi
+
+cleanup_files

.ci/scripts/test_llava.sh

@@ -38,6 +38,7 @@ EXECUTORCH_COMMON_CMAKE_ARGS="                      \
         -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON      \
         -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
         -DEXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=ON \
+        -DEXECUTORCH_BUILD_EXTENSION_NAMED_DATA_MAP=ON \
         -DEXECUTORCH_BUILD_EXTENSION_LLM=ON \
         -DEXECUTORCH_BUILD_EXTENSION_LLM_RUNNER=ON \
         -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON      \

.ci/scripts/test_wheel_package_qnn.sh

@@ -158,10 +158,12 @@ PY
 
   # Install torchao based on the pinned PyTorch version
   "$PIPBIN" install torch=="${TORCH_VERSION}.${NIGHTLY_VERSION}" --index-url "https://download.pytorch.org/whl/nightly/cpu"
+  "$PIPBIN" install wheel
 
   # Install torchao based on the pinned commit from third-party/ao submodule
   pushd "$REPO_ROOT/third-party/ao" > /dev/null
-  USE_CPP=0 "$PYBIN" setup.py develop
+  export USE_CPP=0
+  "$PIPBIN" install . --no-build-isolation
   popd > /dev/null
 
   echo "=== [$LABEL] Import smoke tests ==="

.githooks/README.md

@@ -0,0 +1,57 @@
+# Git Hooks
+
+This directory contains Git hooks for the ExecuTorch repository.
+
+## Pre-commit Hook
+
+The pre-commit hook automatically updates the PyTorch commit pin in `.ci/docker/ci_commit_pins/pytorch.txt` whenever `torch_pin.py` is modified.
+
+### How It Works
+
+1. When you commit changes to `torch_pin.py`, the hook detects the change
+2. It parses the `NIGHTLY_VERSION` field (e.g., `dev20251004`)
+3. Converts it to a date string (e.g., `2025-10-04`)
+4. Fetches the corresponding commit hash from the PyTorch nightly branch at https://github.com/pytorch/pytorch/tree/nightly
+5. Updates `.ci/docker/ci_commit_pins/pytorch.txt` with the new commit hash
+6. Automatically stages the updated file for commit
+
+### Installation
+
+To install the Git hooks, run:
+
+```bash
+.githooks/install.sh
+```
+
+This will copy the pre-commit hook to `.git/hooks/` and make it executable.
+
+### Manual Usage
+
+You can also run the update script manually at any time:
+
+```bash
+python .github/scripts/update_pytorch_pin.py
+```
+
+### Uninstalling
+
+To remove the pre-commit hook:
+
+```bash
+rm .git/hooks/pre-commit
+```
+
+## Troubleshooting
+
+If the hook fails during a commit:
+
+1. Check that Python 3 is available in your PATH
+2. Ensure you have internet connectivity to fetch commits from GitHub
+3. Verify that the `NIGHTLY_VERSION` in `torch_pin.py` is in the correct format (`devYYYYMMDD`)
+4. Make sure the corresponding nightly release exists in the PyTorch nightly branch
+
+You can run the script manually to see detailed error messages:
+
+```bash
+python .github/scripts/update_pytorch_pin.py
+```

.githooks/install.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+
+# Script to install Git hooks from .githooks directory
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+GIT_DIR="$(git rev-parse --git-dir)"
+HOOKS_DIR="${GIT_DIR}/hooks"
+
+echo "Installing Git hooks..."
+
+# Install pre-commit hook
+echo "📦 Installing pre-commit hook..."
+cp "${SCRIPT_DIR}/pre-commit" "${HOOKS_DIR}/pre-commit"
+chmod +x "${HOOKS_DIR}/pre-commit"
+echo "✅ pre-commit hook installed"
+
+echo ""
+echo "🎉 Git hooks installed successfully!"
+echo ""
+echo "The pre-commit hook will automatically update .ci/docker/ci_commit_pins/pytorch.txt"
+echo "whenever you commit changes to torch_pin.py"

.githooks/pre-commit

@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+
+# Pre-commit hook to automatically update PyTorch commit pin and sync c10 directories when torch_pin.py changes
+
+# Check if torch_pin.py is being committed
+if git diff --cached --name-only | grep -q "^torch_pin.py$"; then
+    echo "🔍 Detected changes to torch_pin.py"
+    echo "📝 Updating PyTorch commit pin and syncing c10 directories..."
+
+    # Run the update script (which now also syncs c10 directories)
+    if python .github/scripts/update_pytorch_pin.py; then
+        # Stage any modified files (pytorch.txt and grafted c10 files)
+        if ! git diff --quiet .ci/docker/ci_commit_pins/pytorch.txt; then
+            git add .ci/docker/ci_commit_pins/pytorch.txt
+            echo "📌 Staged .ci/docker/ci_commit_pins/pytorch.txt"
+        fi
+
+        # Stage any grafted c10 files
+        if ! git diff --quiet runtime/core/portable_type/c10/; then
+            git add runtime/core/portable_type/c10/
+            echo "📌 Staged grafted c10 files"
+        fi
+    else
+        echo "❌ Failed to update PyTorch commit pin"
+        echo "Please run: python .github/scripts/update_pytorch_pin.py"
+        exit 1
+    fi
+fi
+
+exit 0