Merge branch 'main' into jz/quantize_kwargs

Author: jackzhxng

.ci/scripts/unittest-linux.sh

@@ -27,20 +27,26 @@ eval "$(conda shell.bash hook)"
 CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
 conda activate "${CONDA_ENV}"
 
-# Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
-source .ci/scripts/setup-vulkan-linux-deps.sh
+if [[ "$BUILD_TOOL" == "cmake" ]]; then
+    # Setup swiftshader and Vulkan SDK which are required to build the Vulkan delegate
+    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" \
-.ci/scripts/setup-linux.sh "$BUILD_TOOL" "$BUILD_MODE"
+    PYTHON_EXECUTABLE=python \
+    EXECUTORCH_BUILD_PYBIND=ON \
+    CMAKE_ARGS="-DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON" \
+    .ci/scripts/setup-linux.sh "$BUILD_TOOL" "$BUILD_MODE"
 
-# Install llama3_2_vision dependencies.
-PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh
+    # Install llama3_2_vision dependencies.
+    PYTHON_EXECUTABLE=python ./examples/models/llama3_2_vision/install_requirements.sh
 
-if [[ "$BUILD_TOOL" == "cmake" ]]; then
     .ci/scripts/unittest-linux-cmake.sh
 elif [[ "$BUILD_TOOL" == "buck2" ]]; then
+    # Removing this breaks sccache in the Buck build, apparently
+    # because TMPDIR gets messed up? Please feel free to fix this and
+    # speed up this CI job!
+    PYTHON_EXECUTABLE=python \
+    .ci/scripts/setup-linux.sh "$BUILD_TOOL" "$BUILD_MODE"
+
     .ci/scripts/unittest-buck2.sh
 else
     echo "Unknown build tool $BUILD_TOOL"

.ci/scripts/unittest-macos.sh

@@ -30,19 +30,19 @@ export TMP_DIR=$(mktemp -d)
 export PATH="${TMP_DIR}:$PATH"
 trap 'rm -rfv ${TMP_DIR}' EXIT
 
-# 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" \
-${CONDA_RUN} --no-capture-output \
-.ci/scripts/setup-macos.sh "${BUILD_TOOL}" "${BUILD_MODE}"
+if [[ "$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" \
+    ${CONDA_RUN} --no-capture-output \
+    .ci/scripts/setup-macos.sh "${BUILD_TOOL}" "${BUILD_MODE}"
 
-# Install llama3_2_vision dependencies.
-PYTHON_EXECUTABLE=python \
-${CONDA_RUN} --no-capture-output \
-./examples/models/llama3_2_vision/install_requirements.sh
+    # Install llama3_2_vision dependencies.
+    PYTHON_EXECUTABLE=python \
+    ${CONDA_RUN} --no-capture-output \
+    ./examples/models/llama3_2_vision/install_requirements.sh
 
-if [[ "$BUILD_TOOL" == "cmake" ]]; then
     .ci/scripts/unittest-macos-cmake.sh
 elif [[ "$BUILD_TOOL" == "buck2" ]]; then
     .ci/scripts/unittest-buck2.sh

.github/workflows/_android.yml

@@ -29,7 +29,7 @@ jobs:
         export ARTIFACTS_DIR_NAME=artifacts-to-be-uploaded
 
         # Build LLM Demo for Android
-        bash build/build_android_llm_demo.sh ${ARTIFACTS_DIR_NAME}
+        bash build/build_android_library.sh ${ARTIFACTS_DIR_NAME}
 
   # Running Android emulator directly on the runner and not using Docker
   run-emulator:

.github/workflows/android-perf.yml

@@ -363,7 +363,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
 
         export ANDROID_ABIS="arm64-v8a"
-        PYTHON_EXECUTABLE=python EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029 bash build/build_android_llm_demo.sh ${ARTIFACTS_DIR_NAME}
+        PYTHON_EXECUTABLE=python EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029 bash build/build_android_library.sh ${ARTIFACTS_DIR_NAME}
 
   # Let's see how expensive this job is, we might want to tone it down by running it periodically
   benchmark-on-device:

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

@@ -53,7 +53,7 @@ jobs:
         export ARTIFACTS_DIR_NAME=artifacts-to-be-uploaded
 
         # Build LLM Demo for Android
-        bash build/build_android_llm_demo.sh ${ARTIFACTS_DIR_NAME}
+        bash build/build_android_library.sh ${ARTIFACTS_DIR_NAME}
 
         shasum -a 256 "${ARTIFACTS_DIR_NAME}/llm_demo/executorch.aar"
 

.github/workflows/pull.yml

@@ -56,6 +56,30 @@ jobs:
         # Build and test ExecuTorch with the add model on portable backend.
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_model.sh "add" "${BUILD_TOOL}" "portable"
 
+  test-pip-install-editable-mode-linux:
+    name: test-pip-install-editable-mode-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-clang12
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+        # Debug
+        which pip
+        PYTHON_EXECUTABLE=python bash ./install_executorch.sh --editable --pybind xnnpack --use-pt-pinned-commit
+        # Try to import extension library
+        python -c "from executorch.extension.llm.custom_ops import custom_ops"
+
   test-models-linux:
     name: test-models-linux
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
@@ -480,7 +504,7 @@ jobs:
 
         # Setup install_requirements for llama
         PYTHON_EXECUTABLE=python bash examples/models/llama/install_requirements.sh
-        
+
         # Test static llama weight sharing and accuracy
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_qnn_static_llama.sh
 

.github/workflows/trunk.yml

@@ -36,6 +36,31 @@ jobs:
 
           PYTHONPATH="${PWD}" python .ci/scripts/gather_test_models.py --target-os macos --event "${GITHUB_EVENT_NAME}"
 
+  test-pip-install-editable-mode-macos:
+    name: test-pip-install-editable-mode-macos
+    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+    with:
+      runner: macos-m1-stable
+      python-version: '3.11'
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+        # Debug
+        which pip
+        bash .ci/scripts/setup-conda.sh
+        PYTHON_EXECUTABLE=python ${CONDA_RUN} bash ./install_executorch.sh --editable --pybind xnnpack
+        # Try to import extension library
+        python -c "from executorch.extension.llm.custom_ops import custom_ops"
+
   test-models-macos:
     name: test-models-macos
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main

backends/arm/arm_vela.py

@@ -12,7 +12,13 @@
 from typing import List
 
 import numpy as np
-from ethosu.vela import vela  # type: ignore
+
+try:
+    from ethosu.vela import vela  # type: ignore
+
+    has_vela = True
+except ImportError:
+    has_vela = False
 
 
 # Pack either input or output tensor block, compose the related arrays into
@@ -45,6 +51,11 @@ def vela_compile(
     """
     Compile a TOSA graph to a binary stream for ArmBackendEthosU using Vela.
     """
+    if not has_vela:
+        raise RuntimeError(
+            "ethos-u-vela pip package couldn't be imported. Make sure it's installed!"
+        )
+
     with tempfile.TemporaryDirectory() as tmpdir:
         tosaname = "out.tosa"
         tosa_path = os.path.join(tmpdir, tosaname)

backends/arm/operators/op_avg_pool2d.py

@@ -41,7 +41,7 @@ def _build_generic_avgpool2d(
         output: TosaArg,
         input_zp: int,
         output_zp: int,
-        accumulator_type,
+        accumulator_type: ts.DType,
     ) -> None:
         input_tensor = inputs[0]
 

backends/arm/operators/op_conv2d.py

@@ -22,8 +22,6 @@
 from executorch.backends.arm.tosa_quant_utils import build_rescale_conv_output
 from executorch.backends.arm.tosa_utils import build_reshape, tosa_shape
 
-from serializer.tosa_serializer import TosaOp
-
 
 @register_node_visitor
 class Conv2dVisitor(NodeVisitor):
@@ -36,8 +34,12 @@ def __init__(self, *args):
     # `(input + 2 * pad - dilation * (weight - 1) - 1) / stride`
     # must be an integer, but tosa currently strictly require this property.
     # This function adjusts the pad value to meet the requirement.
-    def adjust_pad_if_needed(self, input, weight, stride, pad, dilation):
-        mod_remainder = (input + 2 * pad - dilation * (weight - 1) - 1) % stride
+    def adjust_pad_if_needed(
+        self, input_size: int, input_weight: int, stride: int, pad: int, dilation: int
+    ) -> int:
+        mod_remainder = (
+            input_size + 2 * pad - dilation * (input_weight - 1) - 1
+        ) % stride
 
         # No need to adjust
         if mod_remainder == 0:
@@ -143,11 +145,11 @@ def define_node(
             build_reshape(
                 tosa_graph, weight.name, weight_post_shape, weight_reshaped.name
             )
-            tosa_op = TosaOp.Op().DEPTHWISE_CONV2D
+            tosa_op = ts.TosaOp.Op().DEPTHWISE_CONV2D
             weight_name = weight_reshaped.name
         else:
             """Regular convolution case"""
-            tosa_op = TosaOp.Op().CONV2D
+            tosa_op = ts.TosaOp.Op().CONV2D
             weight_name = weight.name
 
         tosa_graph.addOperator(