Merge branch 'main' into fix-et-lowered-module

Author: metascroy

.ci/scripts/export_model_artifact.sh

@@ -34,7 +34,7 @@ Arguments:
 
   output_dir   Output directory for artifacts (optional, default: current directory)
 
-  mode         Export mode (optional, default: auto-detect based on model and device)
+  mode         Export mode (optional, default: vr-streaming)
                Supported modes:
                  - vr-streaming: Voxtral Realtime streaming mode
                  - vr-offline: Voxtral Realtime offline mode
@@ -256,16 +256,9 @@ if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   fi
 
   # Determine streaming mode based on MODE parameter
-  USE_STREAMING="false"
-  if [ "$MODE" = "vr-streaming" ]; then
-    USE_STREAMING="true"
-  elif [ "$MODE" = "vr-offline" ]; then
+  USE_STREAMING="true"
+  if [ "$MODE" = "vr-offline" ]; then
     USE_STREAMING="false"
-  elif [ -z "$MODE" ]; then
-    # Auto-detect: XNNPACK uses streaming, others use offline
-    if [ "$DEVICE" = "xnnpack" ]; then
-      USE_STREAMING="true"
-    fi
   fi
 
   # Configure export and preprocessor based on streaming mode

.ci/scripts/test_lora.sh

@@ -12,7 +12,8 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 cmake_install_executorch_libraries() {
     echo "Installing libexecutorch.a, libextension_module.so, libportable_ops_lib.a"
     rm -rf cmake-out
-    cmake --workflow llm-release
+    cmake --preset llm-release -DEXECUTORCH_ENABLE_LOGGING=ON
+    cmake --build --preset llm-release-install
 }
 
 cmake_build_llama_runner() {

.ci/scripts/test_lora_multimethod.sh

@@ -12,7 +12,8 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 cmake_install_executorch_libraries() {
     echo "Installing libexecutorch.a, libextension_module.so, libportable_ops_lib.a"
     rm -rf cmake-out
-    cmake --workflow llm-release
+    cmake --preset llm-release -DEXECUTORCH_ENABLE_LOGGING=ON
+    cmake --build --preset llm-release-install
 }
 
 cmake_build_llama_runner() {

.ci/scripts/test_model_e2e.sh

@@ -272,16 +272,9 @@ case "$MODEL_NAME" in
   voxtral_realtime)
     RUNNER_ARGS="--model_path ${MODEL_DIR}/model.pte --tokenizer_path ${MODEL_DIR}/$TOKENIZER_FILE --preprocessor_path ${MODEL_DIR}/$PREPROCESSOR --audio_path ${MODEL_DIR}/$AUDIO_FILE --temperature 0"
     # Determine streaming mode based on MODE parameter
-    USE_STREAMING="false"
-    if [ "$MODE" = "vr-streaming" ]; then
-      USE_STREAMING="true"
-    elif [ "$MODE" = "vr-offline" ]; then
+    USE_STREAMING="true"
+    if [ "$MODE" = "vr-offline" ]; then
       USE_STREAMING="false"
-    elif [ -z "$MODE" ]; then
-      # Auto-detect: XNNPACK uses streaming, others use offline
-      if [ "$DEVICE" = "xnnpack" ]; then
-        USE_STREAMING="true"
-      fi
     fi
     # Add streaming flag if needed
     if [ "$USE_STREAMING" = "true" ]; then

.github/workflows/_android.yml

@@ -5,98 +5,89 @@ on:
   workflow_dispatch:
 
 jobs:
-  build-llm-demo:
-    name: build-llm-demo
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+  build-android:
+    name: build-android
+    runs-on: 16-core-ubuntu
     permissions:
       id-token: write
       contents: read
-    with:
-      runner: linux.2xlarge
-      docker-image: ci-image:executorch-ubuntu-22.04-clang12-android
-      submodules: 'recursive'
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      timeout: 90
-      upload-artifact: android-apps
-      upload-artifact-to-s3: true
-      script: |
-        set -eux
+    env:
+      PYTHON_EXECUTABLE: python3
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
 
-        # Use sccache for NDK compiler as well
-        export CMAKE_CXX_COMPILER_LAUNCHER=sccache
-        export CMAKE_C_COMPILER_LAUNCHER=sccache
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
 
-        # 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}"
-        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool buck2
-        export ARTIFACTS_DIR_NAME=artifacts-to-be-uploaded
-        mkdir -p ${ARTIFACTS_DIR_NAME}/
+      - name: Install Python dependencies
+        run: |
+          pip install torch executorch
 
-        # Build LLM Demo for Android
-        export BUILD_AAR_DIR=aar-out
-        mkdir -p $BUILD_AAR_DIR
-        bash scripts/build_android_library.sh
-        cp ${BUILD_AAR_DIR}/executorch.aar $ARTIFACTS_DIR_NAME
+      - name: Set up Android NDK
+        run: |
+          # Use the pre-installed NDK from the GH runner
+          ANDROID_NDK=$(ls -d "$ANDROID_HOME/ndk/"*/ 2>/dev/null | sort -V | tail -1)
+          if [ -z "$ANDROID_NDK" ]; then
+            echo "No Android NDK found at $ANDROID_HOME/ndk/"
+            exit 1
+          fi
+          ANDROID_NDK="${ANDROID_NDK%/}"
+          echo "Using Android NDK: $ANDROID_NDK"
+          echo "ANDROID_NDK=$ANDROID_NDK" >> "$GITHUB_ENV"
+
+      - name: Build Android library
+        run: |
+          set -eux
+          export BUILD_AAR_DIR=aar-out
+          mkdir -p $BUILD_AAR_DIR
+          bash scripts/build_android_library.sh
 
-        mkdir -p ${ARTIFACTS_DIR_NAME}/library_test_dir
-        bash extension/android/executorch_android/android_test_setup.sh
-        (cd extension/android; ANDROID_HOME="${ANDROID_SDK:-/opt/android/sdk}" ./gradlew :executorch_android:assembleAndroidTest)
-        cp extension/android/executorch_android/build/outputs/apk/androidTest/debug/executorch_android-debug-androidTest.apk "${ARTIFACTS_DIR_NAME}/library_test_dir"
+      - name: Build Android test APK
+        run: |
+          set -eux
+          bash extension/android/executorch_android/android_test_setup.sh
+          cd extension/android
+          ANDROID_HOME="${ANDROID_HOME}" ./gradlew :executorch_android:assembleAndroidTest
 
-        mkdir -p ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom
-        bash examples/models/llama/install_requirements.sh
-        bash ".ci/scripts/test_llama.sh" -model stories110M -build_tool cmake -dtype fp16 -mode portable -upload ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: android-apps
+          path: |
+            aar-out/executorch.aar
+            extension/android/executorch_android/build/outputs/apk/androidTest/debug/executorch_android-debug-androidTest.apk
 
   # Running Android emulator directly on the runner and not using Docker
   run-emulator:
-    needs: build-llm-demo
+    needs: build-android
     # NB: This job runs on the shared 8-core-ubuntu runner; KVM / bare-metal assumptions no longer apply
     runs-on: 8-core-ubuntu
     env:
-      ANDROID_NDK_VERSION: r28c
       API_LEVEL: 34
     steps:
       - name: Setup SSH (Click me for login details)
         uses: pytorch/test-infra/.github/actions/setup-ssh@main
         with:
           github-secret: ${{ secrets.GITHUB_TOKEN }}
           instructions: |
-            This is used to run Android emulators, ANDROID_HOME is installed at /opt/android/sdk
+            This is used to run Android emulators
 
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
         with:
           submodules: false
 
-      - name: Setup conda
-        uses: pytorch/test-infra/.github/actions/setup-miniconda@main
+      - name: Download Artifacts
+        uses: actions/download-artifact@v4
         with:
-          python-version: '3.10'
-
-      - name: Install Android dependencies
-        shell: bash
-        run: |
-          set -eux
+          name: android-apps
 
-          # Reuse the script that install Android on ET Docker image
-          sudo -E bash .ci/docker/common/install_android.sh
-
-          # After https://github.com/ReactiveCircus/android-emulator-runner/releases/tag/v2.33.0 release,
-          # it seems that we need to chown the Android setup to the current user instead of root to
-          # avoid permission issue
-          sudo chown -R "${USER}" /opt/android
-
-      - name: Download Artifacts
-        shell: bash
+      - name: Prepare test APK
         run: |
-          set -eux
-          curl -O https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/fp32-xnnpack-custom/model.zip
-          curl -o android-test-debug-androidTest.apk https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/library_test_dir/executorch_android-debug-androidTest.apk
-          unzip model.zip
-          mv *.pte model.pte
-
-      - name: Gradle cache
-        uses: gradle/actions/setup-gradle@v3
+          cp extension/android/executorch_android/build/outputs/apk/androidTest/debug/executorch_android-debug-androidTest.apk android-test-debug-androidTest.apk
 
       - name: AVD cache
         uses: actions/cache@v4
@@ -109,21 +100,15 @@ jobs:
 
       # NB: It takes about 10m to cold boot the emulator here
       - name: Run Android emulator
-        env:
-          ANDROID_HOME: /opt/android/sdk
         uses: reactivecircus/android-emulator-runner@v2
         with:
           api-level: ${{ env.API_LEVEL }}
           arch: x86_64
           script: ./scripts/run_android_emulator.sh
-          # NB: This is to boot the emulator faster following the instructions on
-          # https://github.com/ReactiveCircus/android-emulator-runner. The max number
-          # of cores we can set is 6, any higher number will be reduced to 6.
           cores: 6
           ram-size: 16384M
           heap-size: 12288M
           force-avd-creation: false
           disable-animations: true
           emulator-options: -no-snapshot-save -no-window -gpu swiftshader_indirect -noaudio -no-boot-anim -camera-back none
-          # This is to make sure that the job doesn't fail flakily
           emulator-boot-timeout: 900

.github/workflows/lint.yml

@@ -21,58 +21,58 @@ jobs:
     uses: ./.github/workflows/_get-changed-files.yml
 
   lintrunner-mypy:
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     needs: [get-changed-files]
+    runs-on: ubuntu-latest
     permissions:
-      id-token: write
       contents: read
     if: |
-      github.repository_owner == 'pytorch' && (
-        needs.get-changed-files.outputs.changed-files == '*' ||
-        contains(needs.get-changed-files.outputs.changed-files, '.py') ||
-        contains(needs.get-changed-files.outputs.changed-files, '.pyi')
-      )
-    with:
-      runner: linux.2xlarge
-      docker-image: ci-image:executorch-ubuntu-22.04-linter
-      submodules: 'recursive'
-      fetch-depth: 0
-      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 "${CONDA_ENV}"
-
-        # For mypy linting, we need to first install executorch first so that
-        # it builds the python package information.
-        BUILD_TOOL="cmake"
-        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool "${BUILD_TOOL}"
-
-        CACHE_DIRECTORY="/tmp/.lintbin"
-        # Try to recover the cached binaries
-        if [[ -d "${CACHE_DIRECTORY}" ]]; then
-          # It's ok to fail this as lintrunner init would download these binaries
-          # again if they do not exist
-          cp -r "${CACHE_DIRECTORY}" . || true
-        fi
-
-        RC=0
-        # Run lintrunner on all files. pytorch/pytorch notes that mypy
-        # in particular needs this.
-        if ! lintrunner --force-color --all-files --take MYPY --tee-json=lint.json 2> /dev/null; then
-          echo ""
-          echo -e "\e[1m\e[36mYou can reproduce these results locally by using \`lintrunner --take MYPY\`. (If you don't get the same results, run \'lintrunner init\' to update your local linter)\e[0m"
-          echo -e "\e[1m\e[36mSee https://github.com/pytorch/pytorch/wiki/lintrunner for setup instructions.\e[0m"
-          RC=1
-        fi
-
-        # Use jq to massage the JSON lint output into GitHub Actions workflow commands.
-        jq --raw-output \
-          '"::\(if .severity == "advice" or .severity == "disabled" then "warning" else .severity end) file=\(.path),line=\(.line),col=\(.char),title=\(.code) \(.name)::" + (.description | gsub("\\n"; "%0A"))' \
-          lint.json || true
-
-        exit $RC
+      needs.get-changed-files.outputs.changed-files == '*' ||
+      contains(needs.get-changed-files.outputs.changed-files, '.py') ||
+      contains(needs.get-changed-files.outputs.changed-files, '.pyi')
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+
+      - name: Checkout torchao submodule
+        run: git submodule update --init third-party/ao
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+          cache-dependency-path: requirements-lintrunner.txt
+
+      - name: Install dependencies
+        run: |
+          pip install torch --index-url https://download.pytorch.org/whl/cpu
+          pip install lintrunner==0.12.7 lintrunner-adapters==0.13.0
+          pip install -r requirements-lintrunner.txt
+          USE_CPP=0 pip install --no-build-isolation third-party/ao
+          pip install pytest numpy parameterized huggingface_hub transformers timm expecttest types-requests
+          pip install torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+
+      - name: Generate mypy stubs for C++ bindings
+        run: |
+          cp extension/pybindings/pybindings.pyi extension/pybindings/_portable_lib.pyi
+
+      - name: Run mypy
+        run: |
+          lintrunner init
+          RC=0
+          if ! lintrunner --force-color --all-files --take MYPY --tee-json=lint.json 2> /dev/null; then
+            echo ""
+            echo -e "\e[1m\e[36mYou can reproduce these results locally by using \`lintrunner --take MYPY\`. (If you don't get the same results, run \'lintrunner init\' to update your local linter)\e[0m"
+            echo -e "\e[1m\e[36mSee https://github.com/pytorch/pytorch/wiki/lintrunner for setup instructions.\e[0m"
+            RC=1
+          fi
+
+          # Use jq to massage the JSON lint output into GitHub Actions workflow commands.
+          jq --raw-output \
+            '"::\(if .severity == "advice" or .severity == "disabled" then "warning" else .severity end) file=\(.path),line=\(.line),col=\(.char),title=\(.code) \(.name)::" + (.description | gsub("\\n"; "%0A"))' \
+            lint.json || true
+
+          exit $RC
 
   lintrunner:
     needs: [get-changed-files]

CMakeLists.txt

@@ -987,10 +987,11 @@ if(EXECUTORCH_BUILD_PYBIND)
       -fexceptions>
   )
 
-  # util lib
+  # util lib PyTorch requires C++20, so util must be compiled with C++20.
   add_library(
     util ${CMAKE_CURRENT_SOURCE_DIR}/extension/aten_util/aten_bridge.cpp
   )
+  set_target_properties(util PROPERTIES CXX_STANDARD 20)
   target_include_directories(
     util PUBLIC ${_common_include_directories} ${TORCH_INCLUDE_DIRS}
   )
@@ -1002,8 +1003,11 @@ if(EXECUTORCH_BUILD_PYBIND)
   # pybind portable_lib
   pybind11_add_module(portable_lib SHARED extension/pybindings/pybindings.cpp)
   # The actual output file needs a leading underscore so it can coexist with
-  # portable_lib.py in the same python package.
-  set_target_properties(portable_lib PROPERTIES OUTPUT_NAME "_portable_lib")
+  # portable_lib.py in the same python package. PyTorch requires C++20, so
+  # pybindings must be compiled with C++20.
+  set_target_properties(
+    portable_lib PROPERTIES OUTPUT_NAME "_portable_lib" CXX_STANDARD 20
+  )
   target_compile_definitions(
     portable_lib PUBLIC EXECUTORCH_PYTHON_MODULE_NAME=_portable_lib
   )

CMakePresets.json

@@ -290,13 +290,8 @@
       "name": "arm-ethosu-linux",
       "displayName": "Build ExecuTorch for Arm Ethos-U Linux",
       "inherits": ["common"],
-      "description": "musl declares __assert_fail with int for line; avoid NDEBUG forward-decl mismatch in Release builds",
       "cacheVariables": {
-        "EXECUTORCH_BUILD_ARM_ETHOSU_LINUX": "ON",
-        "EXECUTORCH_BUILD_EXECUTOR_RUNNER": "ON",
-        "EXECUTORCH_BUILD_KERNELS_QUANTIZED": "ON",
-        "CMAKE_C_FLAGS_RELEASE": "-UNDEBUG",
-        "CMAKE_CXX_FLAGS_RELEASE": "-UNDEBUG",
+        "EXECUTORCH_BUILD_PRESET_FILE": "${sourceDir}/tools/cmake/preset/arm_ethosu_linux.cmake",
         "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/examples/arm/ethos-u-setup/aarch64-linux-musl-toolchain.cmake"
       }
     }