Update on "[ET-VK][ez] Fix 8 bit linear compute shader dispatch"

## Context


Currently, for the `q_8w_linear` shader, both the texture and the buffer variants use the same global work group and local work group setting.

Specially, the global work group is set to `{out.numel(), 1, 1}` and the local work group is set to `{64, 1, 1}`.

However, I believe this results in a very poor memory re-use for the texture shader. In this configuration:

* Within a work group each invocation will be requesting a different row of A - 64 rows of A requested in total
* All work groups will be requesting the same row of B
* One work group will load 65 unique rows from A and B

Compare this to a local work group size of `{8, 8, 1}`

* Across the work group, 8 rows will be loaded from A and 8 rows will be loaded from B
* One work group will load 16 unique rows total from A and B

Evidently, there is better memory re-use in the latter work group as fewer unique rows are loaded.

## Changes

Modify the `q_8w_linear` shader to use `{8, 8, 1}` local wg if possible. If `M` is small, then instead use `{4, 16, 1}` or `{2, 32, 1}` to reduce the number of inactive invocations.

Differential Revision: [D71706489](https://our.internmc.facebook.com/intern/diff/D71706489/)

[ghstack-poisoned]

Author: SS-JIA

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-295f2ed4d103017f7e19a7b8263ece606cd629db
+7ae0ce6360b6e4f944906502d20da24c04debee5

.ci/scripts/build_android_instrumentation.sh

@@ -12,30 +12,10 @@ if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
 fi
 which "${PYTHON_EXECUTABLE}"
 
-build_android_test() {
-  mkdir -p extension/android/executorch_android/src/androidTest/resources
-  cp extension/module/test/resources/add.pte extension/android/executorch_android/src/androidTest/resources
-  pushd extension/android
-  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:testDebugUnitTest
-  ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:assembleAndroidTest
-  popd
-}
+mkdir -p "${BUILD_AAR_DIR}"/executorch_android/src/androidTest/resources
+cp extension/module/test/resources/add.pte "${BUILD_AAR_DIR}"/executorch_android/src/androidTest/resources
 
-collect_artifacts_to_be_uploaded() {
-  ARTIFACTS_DIR_NAME="$1"
-  # Collect Java library test
-  JAVA_LIBRARY_TEST_DIR="${ARTIFACTS_DIR_NAME}/library_test_dir"
-  mkdir -p "${JAVA_LIBRARY_TEST_DIR}"
-  cp extension/android/executorch_android/build/outputs/apk/androidTest/debug/*.apk "${JAVA_LIBRARY_TEST_DIR}"
-}
-
-main() {
-  build_android_test
-  if [ -n "$ARTIFACTS_DIR_NAME" ]; then
-    collect_artifacts_to_be_uploaded ${ARTIFACTS_DIR_NAME}
-  fi
-}
-
-if [[ "${BASH_SOURCE[0]}" == "${0}" ]]; then
-  main "$@"
-fi
+pushd "${BUILD_AAR_DIR}"
+ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:testDebugUnitTest
+ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:assembleAndroidTest
+popd

.ci/scripts/gather_test_models.py

@@ -33,7 +33,7 @@
         "dl3": "linux.4xlarge.memory",
         "emformer_join": "linux.4xlarge.memory",
         "emformer_predict": "linux.4xlarge.memory",
-        "phi-4-mini": "linux.4xlarge.memory",
+        "phi_4_mini": "linux.4xlarge.memory",
     }
 }
 

.ci/scripts/test_llama_torchao_lowbit.sh

@@ -78,7 +78,6 @@ ${PYTHON_EXECUTABLE} -m examples.models.llama.export_llama \
     -qmode "torchao:8da${QLINEAR_BITWIDTH}w" \
     --group_size ${QLINEAR_GROUP_SIZE} \
     -E "torchao:${QEMBEDDING_BITWIDTH},${QEMBEDDING_GROUP_SIZE}" \
-    --disable_dynamic_shape \
     -d fp32
 
 # Test run

.ci/scripts/test_model.sh

@@ -100,11 +100,11 @@ test_model() {
       rm "./${MODEL_NAME}.pte"
       return  # Skip running with portable executor runnner since portable doesn't support Qwen's biased linears.
   fi
-  if [[ "${MODEL_NAME}" == "phi-4-mini" ]]; then
+  if [[ "${MODEL_NAME}" == "phi_4_mini" ]]; then
       # Install requirements for export_llama
       bash examples/models/llama/install_requirements.sh
       # Test export_llama script: python3 -m examples.models.llama.export_llama.
-      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/phi-4-mini/config.json
+      "${PYTHON_EXECUTABLE}" -m examples.models.llama.export_llama --model "${MODEL_NAME}" -c examples/models/llama/params/demo_rand_params.pth -p examples/models/phi_4_mini/config.json
       run_portable_executor_runner
       rm "./${MODEL_NAME}.pte"
       return

.ci/scripts/unittest-linux.sh

@@ -21,8 +21,7 @@ if [[ "$BUILD_TOOL" == "cmake" ]]; then
     source .ci/scripts/setup-vulkan-linux-deps.sh
 
     PYTHON_EXECUTABLE=python \
-    EXECUTORCH_BUILD_PYBIND=ON \
-    CMAKE_ARGS="-DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+    CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
     .ci/scripts/setup-linux.sh "$@"
 
     # Install llama3_2_vision dependencies.

.ci/scripts/unittest-macos-buck2.sh

@@ -19,14 +19,13 @@ export TMP_DIR=$(mktemp -d)
 export PATH="${TMP_DIR}:$PATH"
 trap 'rm -rfv ${TMP_DIR}' EXIT
 
-if [[ "$BUILD_TOOL" == "cmake" ]]; then
-    # Setup MacOS dependencies as there is no Docker support on MacOS atm
-    PYTHON_EXECUTABLE=python \
-    EXECUTORCH_BUILD_PYBIND=ON \
-    CMAKE_ARGS="-DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
-    ${CONDA_RUN} --no-capture-output \
-    .ci/scripts/setup-macos.sh "$@"
+# Setup MacOS dependencies as there is no Docker support on MacOS atm
+PYTHON_EXECUTABLE=python \
+CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_COREML=ON -DEXECUTORCH_BUILD_MPS=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+${CONDA_RUN} --no-capture-output \
+.ci/scripts/setup-macos.sh "$@"
 
+if [[ "$BUILD_TOOL" == "cmake" ]]; then
     # Install llama3_2_vision dependencies.
     PYTHON_EXECUTABLE=python \
     ${CONDA_RUN} --no-capture-output \

.ci/scripts/unittest-macos.sh

@@ -60,12 +60,46 @@ install_pytorch_and_domains() {
   # Fetch the target commit
   pushd pytorch || return
   git checkout "${TORCH_VERSION}"
-  git submodule update --init --recursive
 
-  export USE_DISTRIBUTED=1
-  # Then build and install PyTorch
-  python setup.py bdist_wheel
-  pip install "$(echo dist/*.whl)"
+  local system_name=$(uname)
+  if [[ "${system_name}" == "Darwin" ]]; then
+    local platform=$(python -c 'import sysconfig; import platform; v=platform.mac_ver()[0].split(".")[0]; platform=sysconfig.get_platform().split("-"); platform[1]=f"{v}_0"; print("_".join(platform))')
+  fi
+  local python_version=$(python -c 'import platform; v=platform.python_version_tuple(); print(f"{v[0]}{v[1]}")')
+  local torch_release=$(cat version.txt)
+  local torch_short_hash=${TORCH_VERSION:0:7}
+  local torch_wheel_path="cached_artifacts/pytorch/executorch/pytorch_wheels/${system_name}/${python_version}"
+  local torch_wheel_name="torch-${torch_release}%2Bgit${torch_short_hash}-cp${python_version}-cp${python_version}-${platform:-}.whl"
+
+  local cached_torch_wheel="https://gha-artifacts.s3.us-east-1.amazonaws.com/${torch_wheel_path}/${torch_wheel_name}"
+  # Cache PyTorch wheel is only needed on MacOS, Linux CI already has this as part
+  # of the Docker image
+  local torch_wheel_not_found=0
+  if [[ "${system_name}" == "Darwin" ]]; then
+    pip install "${cached_torch_wheel}" || torch_wheel_not_found=1
+  else
+    torch_wheel_not_found=1
+  fi
+
+  # Found no such wheel, we will build it from source then
+  if [[ "${torch_wheel_not_found}" == "1" ]]; then
+    echo "No cached wheel found, continue with building PyTorch at ${TORCH_VERSION}"
+
+    git submodule update --init --recursive
+    USE_DISTRIBUTED=1 python setup.py bdist_wheel
+    pip install "$(echo dist/*.whl)"
+
+    # Only AWS runners have access to S3
+    if command -v aws && [[ -z "${GITHUB_RUNNER:-}" ]]; then
+      for wheel_path in dist/*.whl; do
+        local wheel_name=$(basename "${wheel_path}")
+        echo "Caching ${wheel_name}"
+        aws s3 cp "${wheel_path}" "s3://gha-artifacts/${torch_wheel_path}/${wheel_name}"
+      done
+    fi
+  else
+    echo "Use cached wheel at ${cached_torch_wheel}"
+  fi
 
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt)

.ci/scripts/utils.sh

@@ -8,13 +8,10 @@
 # should typically only contain shell variable assignments. Be sure to export
 # any variables so that subprocesses will see them.
 
-# Enable pybindings so that users can execute ExecuTorch programs from python.
-export EXECUTORCH_BUILD_PYBIND=1
-
 # Ensure that CMAKE_ARGS is defined before referencing it. Defaults to empty
 # if not defined.
 export CMAKE_ARGS="${CMAKE_ARGS:-}"
 
 # Link the XNNPACK backend into the pybindings runtime so that users can execute
 # ExecuTorch programs that delegate to it.
-CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_XNNPACK=ON"
+CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON"