Merge remote-tracking branch 'origin/main' into cache-position-llava

Author: kirklandsign

.ci/scripts/test_wheel_package_qnn.sh

@@ -0,0 +1,201 @@
+#!/bin/bash
+# === CI Wheel Build & Test Script ===
+
+# Exit immediately on error, print each command, and capture all output to build.log
+set -e
+set -x
+exec > >(tee -i build.log) 2>&1
+
+# Save repo root
+REPO_ROOT=$(pwd)
+
+# ----------------------------
+# Dynamically create script_qnn_wheel_test.py
+# ----------------------------
+cat > "/tmp/script_qnn_wheel_test.py" << 'EOF'
+# pyre-ignore-all-errors
+import argparse
+
+import torch
+from executorch.backends.qualcomm.quantizer.quantizer import QnnQuantizer
+from executorch.backends.qualcomm.utils.utils import (
+    generate_htp_compiler_spec,
+    generate_qnn_executorch_compiler_spec,
+    get_soc_to_chipset_map,
+    to_edge_transform_and_lower_to_qnn,
+)
+from executorch.exir.backend.utils import format_delegated_graph
+from executorch.examples.models.model_factory import EagerModelFactory
+from executorch.exir.capture._config import ExecutorchBackendConfig
+from executorch.extension.export_util.utils import save_pte_program
+from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e, prepare_qat_pt2e
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-f", "--output_folder", type=str, default="", help="The folder to store the exported program")
+    parser.add_argument("--soc", type=str, default="SM8650", help="Specify the SoC model.")
+    parser.add_argument("-q", "--quantization", choices=["ptq", "qat"], help="Run post-traininig quantization.")
+    args = parser.parse_args()
+
+    class LinearModule(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.linear = torch.nn.Linear(3, 3)
+        def forward(self, arg):
+            return self.linear(arg)
+        def get_example_inputs(self):
+            return (torch.randn(3, 3),)
+
+    model = LinearModule()
+    example_inputs = model.get_example_inputs()
+
+    if args.quantization:
+        quantizer = QnnQuantizer()
+        m = torch.export.export(model.eval(), example_inputs, strict=True).module()
+        if args.quantization == "qat":
+            m = prepare_qat_pt2e(m, quantizer)
+            m(*example_inputs)
+        elif args.quantization == "ptq":
+            m = prepare_pt2e(m, quantizer)
+            m(*example_inputs)
+        m = convert_pt2e(m)
+    else:
+        m = model
+
+    use_fp16 = True if args.quantization is None else False
+    backend_options = generate_htp_compiler_spec(use_fp16=use_fp16)
+    compile_spec = generate_qnn_executorch_compiler_spec(
+        soc_model=get_soc_to_chipset_map()[args.soc],
+        backend_options=backend_options,
+    )
+    delegated_program = to_edge_transform_and_lower_to_qnn(m, example_inputs, compile_spec)
+    output_graph = format_delegated_graph(delegated_program.exported_program().graph_module)
+    # Ensure QnnBackend is in the output graph
+    assert "QnnBackend" in output_graph
+    executorch_program = delegated_program.to_executorch(
+        config=ExecutorchBackendConfig(extract_delegate_segments=False)
+    )
+    save_pte_program(executorch_program, "linear", args.output_folder)
+
+if __name__ == "__main__":
+    main()
+EOF
+
+# ----------------------------
+# Wheel build and .so checks
+# ----------------------------
+echo "=== Building Wheel Package ==="
+source .ci/scripts/utils.sh
+install_executorch
+EXECUTORCH_BUILDING_WHEEL=1 python setup.py bdist_wheel
+unset EXECUTORCH_BUILDING_WHEEL
+
+WHEEL_FILE=$(ls dist/*.whl | head -n 1)
+echo "Found wheel: $WHEEL_FILE"
+
+PYTHON_VERSION=$1
+# ----------------------------
+# Check wheel does NOT contain qualcomm/sdk
+# ----------------------------
+echo "Checking wheel does not contain qualcomm/sdk..."
+SDK_FILES=$(unzip -l "$WHEEL_FILE" | awk '{print $4}' | grep "executorch/backends/qualcomm/sdk" || true)
+if [ -n "$SDK_FILES" ]; then
+    echo "ERROR: Wheel package contains unexpected qualcomm/sdk files:"
+    echo "$SDK_FILES"
+    exit 1
+else
+    echo "OK: No qualcomm/sdk files found in wheel"
+fi
+
+# ----------------------------
+# Check .so files in the wheel
+# ----------------------------
+echo "Checking for .so files inside the wheel..."
+WHEEL_SO_FILES=$(unzip -l "$WHEEL_FILE" | awk '{print $4}' | grep "executorch/backends/qualcomm/python" || true)
+if [ -z "$WHEEL_SO_FILES" ]; then
+    echo "ERROR: No .so files found in wheel under executorch/backends/qualcomm/python"
+    exit 1
+else
+    echo "Wheel contains the following .so files:"
+    echo "$WHEEL_SO_FILES"
+fi
+
+# ----------------------------
+# Helpers
+# ----------------------------
+get_site_packages_dir () {
+  local PYBIN="$1"
+  "$PYBIN" - <<'PY'
+import sysconfig, sys
+print(sysconfig.get_paths().get("purelib") or sysconfig.get_paths().get("platlib"))
+PY
+}
+
+run_core_tests () {
+  local PYBIN="$1"      # path to python
+  local PIPBIN="$2"     # path to pip
+  local LABEL="$3"      # label to print (conda/venv)
+
+  echo "=== [$LABEL] Installing wheel & deps ==="
+  "$PIPBIN" install --upgrade pip
+  "$PIPBIN" install "$WHEEL_FILE"
+  "$PIPBIN" install torch=="2.9.0.dev20250906" --index-url "https://download.pytorch.org/whl/nightly/cpu"
+  "$PIPBIN" install --pre torchao --index-url "https://download.pytorch.org/whl/nightly/cpu"
+
+  echo "=== [$LABEL] Import smoke tests ==="
+  "$PYBIN" -c "import executorch; print('executorch imported successfully')"
+  "$PYBIN" -c "import executorch.backends.qualcomm; print('executorch.backends.qualcomm imported successfully')"
+
+  echo "=== [$LABEL] List installed executorch/backends/qualcomm/python ==="
+  local SITE_DIR
+  SITE_DIR="$(get_site_packages_dir "$PYBIN")"
+  local SO_DIR="$SITE_DIR/executorch/backends/qualcomm/python"
+  ls -l "$SO_DIR" || echo "Folder does not exist!"
+
+  echo "=== [$LABEL] Run export script to generate linear.pte ==="
+  (cd "$REPO_ROOT" && "$PYBIN" "/tmp/script_qnn_wheel_test.py")
+
+  if [ -f "$REPO_ROOT/linear.pte" ]; then
+      echo "[$LABEL] Model file linear.pte successfully created"
+  else
+      echo "ERROR: [$LABEL] Model file linear.pte was not created"
+      exit 1
+  fi
+}
+
+# ----------------------------
+# Conda environment setup & tests
+# ----------------------------
+echo "=== Testing in Conda env ==="
+TEMP_ENV_DIR=$(mktemp -d)
+echo "Using temporary directory for conda: $TEMP_ENV_DIR"
+conda create -y -p "$TEMP_ENV_DIR/env" python=$PYTHON_VERSION
+# derive python/pip paths inside the conda env
+CONDA_PY="$TEMP_ENV_DIR/env/bin/python"
+CONDA_PIP="$TEMP_ENV_DIR/env/bin/pip"
+# Some images require conda run; keep pip/python direct to simplify path math
+run_core_tests "$CONDA_PY" "$CONDA_PIP" "conda"
+
+# Cleanup conda env
+conda env remove -p "$TEMP_ENV_DIR/env" -y || true
+rm -rf "$TEMP_ENV_DIR"
+
+# ----------------------------
+# Python venv setup & tests
+# ----------------------------
+echo "=== Testing in Python venv ==="
+TEMP_VENV_DIR=$(mktemp -d)
+echo "Using temporary directory for venv: $TEMP_VENV_DIR"
+python3 -m venv "$TEMP_VENV_DIR/venv"
+VENV_PY="$TEMP_VENV_DIR/venv/bin/python"
+VENV_PIP="$TEMP_VENV_DIR/venv/bin/pip"
+
+# Ensure venv has wheel/build basics if needed
+"$VENV_PIP" install --upgrade pip
+
+run_core_tests "$VENV_PY" "$VENV_PIP" "venv"
+
+# Cleanup venv
+rm -rf "$TEMP_VENV_DIR"
+
+echo "=== All tests completed! ==="

.ci/scripts/unittest-buck2.sh

@@ -9,7 +9,11 @@ set -eux
 # TODO: expand this to //...
 # 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/... + \
+# TODO: Make //backends/arm tests use runtime wrapper so we can just query //backends/arm/...
+buck2 query "//backends/apple/... + //backends/arm: + //backends/arm/debug/... + \
+//backends/arm/operator_support/... + //backends/arm/operators/... + \
+//backends/arm/_passes/... + //backends/arm/runtime/... + //backends/arm/tosa/... \
++ //backends/example/... + \
 //backends/mediatek/... + //backends/transforms/... + \
 //backends/xnnpack/... + //configurations/... + //extension/flat_tensor: + \
 //extension/llm/runner: + //kernels/aten/... + //kernels/optimized/... + \

.ci/scripts/wheel/pre_build_script.sh

@@ -9,6 +9,18 @@ set -euxo pipefail
 
 # This script is run before building ExecuTorch binaries
 
+if [[ "$(uname -m)" == "aarch64" ]]; then
+  # On some Linux aarch64 systems, the "atomic" library is not found during linking.
+  # To work around this, replace "atomic" with the literal ${ATOMIC_LIB} so the
+  # build system uses the full path to the atomic library.
+  file="extension/llm/tokenizers/third-party/sentencepiece/src/CMakeLists.txt"
+  sed 's/list(APPEND SPM_LIBS "atomic")/list(APPEND SPM_LIBS ${ATOMIC_LIB})/' \
+    "$file" > "${file}.tmp" && mv "${file}.tmp" "$file"
+
+  grep -n 'list(APPEND SPM_LIBS ${ATOMIC_LIB})' "$file" && \
+    echo "the file $file has been modified for atomic to use full path"
+fi
+
 # Clone nested submodules for tokenizers - this is a workaround for recursive
 # submodule clone failing due to path length limitations on Windows. Eventually,
 # we should update the core job in test-infra to enable long paths before

.ci/scripts/wheel/test_linux_aarch64.py

@@ -0,0 +1,22 @@
+#!/usr/bin/env python
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import test_base
+from examples.models import Backend, Model
+
+if __name__ == "__main__":
+    # coremltools does not support linux aarch64 yet and install from the source fails on runtime
+    # https://github.com/apple/coremltools/issues/1254
+    # https://github.com/apple/coremltools/issues/2195
+    test_base.run_tests(
+        model_tests=[
+            test_base.ModelTest(
+                model=Model.Mv3,
+                backend=Backend.XnnpackQuantizationDelegation,
+            ),
+        ]
+    )

.github/workflows/_get-changed-files.yml

@@ -0,0 +1,43 @@
+name: Get Changed Files
+
+on:
+  workflow_call:
+    outputs:
+      changed-files:
+        description: "List of changed files (space-separated) or '*' if not in a PR"
+        value: ${{ jobs.get-changed-files.outputs.changed-files }}
+
+jobs:
+  get-changed-files:
+    runs-on: ubuntu-latest
+    outputs:
+      changed-files: ${{ steps.get-files.outputs.changed-files }}
+
+    steps:
+      - name: Get changed files
+        id: get-files
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          # Check if we're in a pull request context
+          if [ "${{ github.event_name }}" = "pull_request" ] || [ "${{ github.event_name }}" = "pull_request_target" ]; then
+            echo "Running in PR context"
+
+            # Get the PR number from the github context
+            PR_NUMBER="${{ github.event.number }}"
+
+            # Use gh CLI to get changed files in the PR with explicit repo
+            CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
+
+            if [ -z "$CHANGED_FILES" ]; then
+              echo "No changed files found, setting to '*'"
+              CHANGED_FILES="*"
+            fi
+
+            echo "Changed files: $CHANGED_FILES"
+            echo "changed-files=$CHANGED_FILES" >> "$GITHUB_OUTPUT"
+
+          else
+            echo "Not in PR context, setting changed files to '*'"
+            echo "changed-files=*" >> "$GITHUB_OUTPUT"
+          fi

.github/workflows/apple.yml

@@ -17,7 +17,6 @@ on:
       - scripts/build_apple_llm_demo.sh
       - scripts/create_frameworks.sh
       - .ci/scripts/test_ios_ci.sh
-      - examples/demo-apps/apple_ios/**
       - extension/apple/**
       - extension/benchmark/apple/**
       - extension/module/**

.github/workflows/build-wheels-aarch64-linux.yml

@@ -0,0 +1,66 @@
+# From https://github.com/pytorch/test-infra/wiki/Using-Nova-Reusable-Build-Workflows
+name: Build Aarch64 Linux Wheels
+
+on:
+  pull_request:
+    paths:
+      - .ci/**/*
+      - .github/workflows/build-wheels-aarch64-linux.yml
+      - examples/**/*
+      - pyproject.toml
+      - setup.py
+    tags:
+      - ciflow/binaries/*
+  push:
+    branches:
+      - nightly
+      - release/*
+    tags:
+      # NOTE: Binary build pipelines should only get triggered on release candidate builds
+      # Release candidate tags look like: v1.11.0-rc1
+      - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
+      - ciflow/binaries/*
+  workflow_dispatch:
+
+jobs:
+  generate-matrix:
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    with:
+      package-type: wheel
+      os: linux-aarch64
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      with-cuda: disabled
+      with-rocm: disabled
+      python-versions: '["3.10", "3.11", "3.12"]'
+
+  build:
+    needs: generate-matrix
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - repository: pytorch/executorch
+            pre-script: .ci/scripts/wheel/pre_build_script.sh
+            post-script: .ci/scripts/wheel/post_build_script.sh
+            smoke-test-script: .ci/scripts/wheel/test_linux_aarch64.py
+            package-name: executorch
+    name: ${{ matrix.repository }}
+    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main
+    with:
+      repository: ${{ matrix.repository }}
+      ref: ""
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
+      submodules: recursive
+      env-var-script: .ci/scripts/wheel/envvar_linux.sh
+      pre-script: ${{ matrix.pre-script }}
+      post-script: ${{ matrix.post-script }}
+      package-name: ${{ matrix.package-name }}
+      smoke-test-script: ${{ matrix.smoke-test-script }}
+      trigger-event: ${{ github.event_name }}
+      architecture: aarch64