Add baremetal RISC-V smoke tests (rv32, rv64)

.ci/docker/build.sh

@@ -89,6 +89,11 @@ case "${IMAGE_NAME}" in
     OS_VERSION=24.04
     GCC_VERSION=14
     ;;
+  executorch-ubuntu-26.04-gcc15)
+    LINTRUNNER=""
+    OS_VERSION=26.04
+    GCC_VERSION=15
+    ;;
   *)
     echo "Invalid image name ${IMAGE_NAME}"
     exit 1

.ci/docker/common/install_docs_reqs.sh

@@ -15,8 +15,8 @@ if [ -n "$BUILD_DOCS" ]; then
   curl --retry 3 --retry-all-errors -sL https://deb.nodesource.com/setup_16.x | sudo -E bash -
   sudo apt-get install -y nodejs
 
-  curl --retry 3 --retry-all-errors -sS https://dl.yarnpkg.com/debian/pubkey.gpg | sudo apt-key add -
-  echo "deb https://dl.yarnpkg.com/debian/ stable main" | sudo tee /etc/apt/sources.list.d/yarn.list
+  curl --retry 3 --retry-all-errors -sS https://dl.yarnpkg.com/debian/pubkey.gpg | sudo gpg --dearmor -o /usr/share/keyrings/yarn-archive-keyring.gpg
+  echo "deb [signed-by=/usr/share/keyrings/yarn-archive-keyring.gpg] https://dl.yarnpkg.com/debian/ stable main" | sudo tee /etc/apt/sources.list.d/yarn.list
 
   apt-get update
   apt-get install -y --no-install-recommends yarn

.ci/docker/common/install_linter.sh

@@ -13,7 +13,3 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 # NB: Install all linter dependencies, the caching of lintrunner init could be
 # done after Executorch becomes public
 pip_install -r requirements-lintrunner.txt
-
-# Install google-java-format
-curl -L --retry 3 --retry-all-errors https://github.com/google/google-java-format/releases/download/v1.23.0/google-java-format_linux-x86-64 > /opt/google-java-format
-chmod +x /opt/google-java-format

.ci/scripts/test_riscv_qemu.sh

@@ -4,50 +4,65 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-# CI wrapper: install RISC-V cross-compile + qemu-user tooling, then run the
-# RISC-V smoke test (export, cross-compile, qemu-user execution) via
-# examples/riscv/run.sh. The bundled-IO comparison and Test_result: PASS
-# check are done by run.sh.
+# CI wrapper: install riscv32/64 cross-compile + qemu tooling, then drive
+# examples/riscv/run.sh which does the export, cross-compile, qemu run, and
+# bundled-IO PASS check.
 
 set -eu
 
 script_dir=$(realpath "$(dirname "${BASH_SOURCE[0]}")")
 et_root_dir=$(realpath "${script_dir}/../..")
 
 model="add"
-xnnpack=false
+backend="portable"
 quantize=false
+os="linux"
+arch="rv64"
+qemu_cpu_ext=""
 verbose_xnnpack=false
 debug_xnnpack=false
+build_dir=
 
 usage() {
     cat <<EOF
 Usage: $(basename "$0") [options]
 Options:
-  --model=<NAME>     Which model to export and run (default: add)
-  --xnnpack          Enable the XNNPACK backend (AOT partitioner + runtime)
-  --quantize         Produce an 8-bit quantized model
-  --verbose-xnnpack  Build XNNPACK with XNN_LOG_LEVEL=4 to log microkernel dispatch
-  --debug-xnnpack    Enable XNNPACK partitioner DEBUG logging and dump the lowered graph
-  -h, --help         Show this help
+  --model=<NAME>          Which model to export and run (default: ${model})
+  --quantize              Produce an 8-bit quantized model
+  --backend=<NAME>        AOT backend (portable|xnnpack) (default: ${backend})
+  --os=<NAME>             Target OS (linux|baremetal) (default: ${os})
+  --arch=<NAME>           Target arch (rv32|rv64) (default: ${arch})
+  --qemu-cpu-ext=<EXT>    QEMU -cpu extensions (no rv32/rv64 prefix, default: none)
+  --build-dir=<DIR>       Build/output directory for this configuration (required)
+  --verbose-xnnpack       Build XNNPACK with XNN_LOG_LEVEL=4 to log microkernel dispatch
+  --debug-xnnpack         Enable XNNPACK partitioner DEBUG logging and dump the lowered graph
+  -h, --help              Show this help
 EOF
 }
 
 for arg in "$@"; do
     case $arg in
         --model=*) model="${arg#*=}" ;;
-        --xnnpack) xnnpack=true ;;
         --quantize) quantize=true ;;
+        --backend=*) backend="${arg#*=}" ;;
+        --os=*) os="${arg#*=}" ;;
+        --arch=*) arch="${arg#*=}" ;;
+        --qemu-cpu-ext=*) qemu_cpu_ext="${arg#*=}" ;;
+        --build-dir=*) build_dir="${arg#*=}" ;;
         --debug-xnnpack) debug_xnnpack=true ;;
         --verbose-xnnpack) verbose_xnnpack=true ;;
         -h|--help) usage; exit 0 ;;
         *) echo "Unknown option: $arg" >&2; usage; exit 1 ;;
     esac
 done
 
+if [[ -z "${build_dir}" ]]; then
+    echo "[test_riscv_qemu.sh] --build-dir is required" >&2; usage; exit 1
+fi
+
 run_extra_args=()
-if ${xnnpack}; then
-    run_extra_args+=(--xnnpack)
+if [ -n "${qemu_cpu_ext}" ]; then
+    run_extra_args+=(--qemu-cpu-ext="${qemu_cpu_ext}")
 fi
 if ${quantize}; then
     run_extra_args+=(--quantize)
@@ -59,5 +74,8 @@ if ${verbose_xnnpack}; then
     run_extra_args+=(--verbose-xnnpack)
 fi
 
-bash "${et_root_dir}/examples/riscv/setup.sh"
-bash "${et_root_dir}/examples/riscv/run.sh" --model="${model}" "${run_extra_args[@]}"
+bash "${et_root_dir}/examples/riscv/setup-${os}.sh"
+bash "${et_root_dir}/examples/riscv/run.sh" \
+    --model="${model}" --backend="${backend}" --os="${os}" --arch="${arch}" \
+    --build-dir="${build_dir}" \
+    "${run_extra_args[@]}"

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

@@ -12,8 +12,19 @@ set -eux
 export TORCHINDUCTOR_CACHE_DIR="$(mktemp -d "${RUNNER_TEMP:-/tmp}/torchinductor_cache_XXXXXX")"
 trap 'rm -rf "${TORCHINDUCTOR_CACHE_DIR}"' EXIT
 
-# Run pytest with coverage
-${CONDA_RUN} pytest -n auto --cov=./ --cov-report=xml
+# TODO(SS-JIA): AOTI tests hang on macOS CI runners — the thread blocks in
+# native C/C++ code (dlopen / inductor compilation) so faulthandler cannot
+# even produce a traceback. Diagnosis ongoing in #19886.
+AOTI_SKIPS=(
+  --ignore=examples/models/llama3_2_vision/preprocess/test_preprocess.py
+  --ignore=examples/models/llama3_2_vision/vision_encoder/test/test_vision_encoder.py
+  --ignore=examples/models/llama3_2_vision/text_decoder/test/test_text_decoder.py
+  --deselect=extension/llm/modules/test/test_position_embeddings.py::TilePositionalEmbeddingTest::test_tile_positional_embedding_aoti
+  --deselect=extension/llm/modules/test/test_position_embeddings.py::TiledTokenPositionalEmbeddingTest::test_tiled_token_positional_embedding_aoti
+  --deselect=extension/llm/modules/test/test_attention.py::AttentionTest::test_attention_aoti
+)
+
+${CONDA_RUN} pytest -n auto --cov=./ --cov-report=xml "${AOTI_SKIPS[@]}"
 # Run gtest
 LLVM_PROFDATA="xcrun llvm-profdata" LLVM_COV="xcrun llvm-cov" \
 ${CONDA_RUN} test/run_oss_cpp_tests.sh

.github/scripts/propose_ghstack_orig_pr.py

@@ -52,12 +52,9 @@ def extract_stack_from_body(pr_body: str) -> List[int]:
     """
 
     prs = []
-    ghstack_begin = (
-        "Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):"
-    )
     ghstack_begin_seen = False
     for line in pr_body.splitlines():
-        if ghstack_begin in line:
+        if line.startswith("Stack from [ghstack]"):
             ghstack_begin_seen = True
         if not ghstack_begin_seen:
             continue

.github/workflows/_test_riscv.yml

@@ -13,35 +13,44 @@ on:
         type: number
         default: 30
       model:
-        description: 'Which model to run. Possible values are: add, mv2 (mobilenetv2)'
+        description: 'Which model to run (add, mv2, mobilebert, llama2, resnet18, yolo26)'
         required: false
         type: string
         default: 'add'
-      xnnpack:
-        description: 'Whether to enable XNNPACK'
-        required: false
-        type: boolean
-        default: false
       quantize:
         description: 'Produce an 8-bit quantized model'
         required: false
         type: boolean
         default: false
-      qemu-cpu:
-        description: 'Configuration(s) for the CPU to emulate with QEMU, expecting a JSON array'
-        required: true
+      backend:
+        description: 'AOT backend to lower to (portable|xnnpack)'
+        required: false
         type: string
-      docker-image:
-        description: 'The docker image to use for this job'
+        default: 'portable'
+      os:
+        description: 'Target OS for the runner (linux|baremetal)'
         required: false
         type: string
+        default: 'linux'
+      arch:
+        description: 'Target architecture (rv32|rv64)'
+        required: false
+        type: string
+        default: 'rv64'
+      qemu-cpu-ext:
+        description: >-
+          JSON array of QEMU -cpu *extension* strings (no rv32/rv64 prefix).
+          The script splices each entry with `arch` to form the final -cpu
+          value. Use [""] for plain base-ISA runs.
+        required: true
+        type: string
 
 jobs:
   run:
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
       runner: linux.2xlarge
-      docker-image: ci-image:executorch-ubuntu-24.04-gcc14
+      docker-image: ${{ inputs.os == 'linux' && 'ci-image:executorch-ubuntu-24.04-gcc14' || 'ci-image:executorch-ubuntu-26.04-gcc15' }}
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: ${{ inputs.timeout }}
@@ -55,20 +64,26 @@ jobs:
         # Allows failure in `echo | jq | while read` pipeline to bubble up and fail the workflow
         set -o pipefail
 
-        echo '${{ inputs.qemu-cpu }}' | jq -r '.[]' | while IFS= read -r qemu_cpu; do
-          export QEMU_CPU="${qemu_cpu}"
-          export GCC_VERSION=14
+        echo '${{ inputs.qemu-cpu-ext }}' | jq -r '.[]' | while IFS= read -r qemu_cpu_ext; do
+          variant_slug="${qemu_cpu_ext//,/_}"; variant_slug="${variant_slug//=/_}"; variant_slug="${variant_slug:-base}"
+          build_dir="riscv_test/${{ inputs.model }}${{ inputs.quantize && '_q' || '' }}/${{ inputs.backend }}/${{ inputs.os }}-${{ inputs.arch }}-${variant_slug}"
+
           bash .ci/scripts/test_riscv_qemu.sh \
             --model="${{ inputs.model }}" \
-            ${{ inputs.xnnpack && '--xnnpack --verbose-xnnpack' || '' }} \
+            --backend="${{ inputs.backend }}" \
+            --os="${{ inputs.os }}" \
+            --arch="${{ inputs.arch }}" \
+            --qemu-cpu-ext="${qemu_cpu_ext}" \
+            --build-dir="${build_dir}" \
+            ${{ inputs.backend == 'xnnpack' && '--verbose-xnnpack' || '' }} \
             ${{ inputs.quantize && '--quantize' || '' }}
 
-          # We only generate riscv_test/${{ inputs.model }}_riscv.etdump.json from `--verbose-xnnpack`.
-          if ${{ inputs.xnnpack }}; then
-            # Generate markdown table from riscv_test/${{ inputs.model }}_riscv.etdump.json, sorted by sum_ms
+          # We only generate run.etdump.json from `--verbose-xnnpack`.
+          if [[ "${{ inputs.backend }}" == "xnnpack" ]]; then
+            # Generate markdown table from ${build_dir}/run.etdump.json, sorted by sum_ms
             (
-              etdump_json="riscv_test/${{ inputs.model }}_riscv.etdump.json"
-              echo "### Model=${{ inputs.model }} XNNPACK=${{ inputs.xnnpack }} Quantize=${{ inputs.quantize }} QEMU_CPU='${QEMU_CPU}'"
+              etdump_json="${build_dir}/run.etdump.json"
+              echo "### Model=${{ inputs.model }} Quantize=${{ inputs.quantize }} Backend=${{ inputs.backend }} OS=${{ inputs.os }} Arch=${{ inputs.arch }}${qemu_cpu_ext:+,${qemu_cpu_ext}}"
               jq -r '
                 def r3: (. * 1000 | round) / 1000;
                 ["Section","Op","Count","Sum (ms)","Avg (ms)","Max (ms)","Microkernels"],

.github/workflows/_unittest.yml

@@ -49,6 +49,7 @@ jobs:
       python-version: '3.11'
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
       script: |
         set -eux
         # This is needed to get the prebuilt PyTorch wheel from S3

.github/workflows/docker-builds.yml

@@ -43,6 +43,7 @@ jobs:
           executorch-ubuntu-22.04-mediatek-sdk,
           executorch-ubuntu-22.04-clang12-android,
           executorch-ubuntu-24.04-gcc14,
+          executorch-ubuntu-26.04-gcc15,
         ]
         include:
           - docker-image-name: executorch-ubuntu-22.04-gcc11-aarch64

.github/workflows/lint.yml

@@ -125,49 +125,3 @@ jobs:
     uses: ./.github/workflows/_link_check.yml
     with:
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-
-  android-java-format:
-    runs-on: ubuntu-latest
-    permissions:
-      contents: read
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-
-      - uses: actions/setup-java@v4
-        with:
-          distribution: 'temurin'
-          java-version: '17'
-
-      - name: Check Java formatting
-        run: |
-          GOOGLE_JAVA_FORMAT_VERSION="1.24.0"
-          curl -sSfL "https://github.com/google/google-java-format/releases/download/v${GOOGLE_JAVA_FORMAT_VERSION}/google-java-format-${GOOGLE_JAVA_FORMAT_VERSION}-all-deps.jar" \
-            -o /tmp/google-java-format.jar
-
-          FILES_NEEDS_FORMAT=$(find extension/android/executorch_android/src/main/java/org/pytorch/executorch \
-                              extension/android/executorch_android/src/main/java/org/pytorch/executorch/extension/llm \
-                              extension/android/executorch_android/src/main/java/org/pytorch/executorch/annotations \
-                              extension/android/executorch_android/src/androidTest/java/org/pytorch/executorch \
-                              extension/benchmark/android/benchmark/app/src/main/java/org/pytorch/minibench \
-                              extension/benchmark/android/benchmark/app/src/androidTest/java/org/pytorch/minibench \
-                              -type f -name "*.java" 2>/dev/null | \
-                              xargs -r java -jar /tmp/google-java-format.jar -n)
-
-          if [ -n "$FILES_NEEDS_FORMAT" ]; then
-            echo "Warning: The following files need formatting:"
-            echo "$FILES_NEEDS_FORMAT"
-            echo ""
-            echo "Please use google-java-format from https://github.com/google/google-java-format/releases/"
-            echo ""
-            echo "To fix, run one of these commands:"
-            echo "  # Using xargs (recommended):"
-            echo "  find <paths> -type f -name '*.java' | xargs google-java-format -i"
-            echo ""
-            echo "  # Or format specific files:"
-            echo "$FILES_NEEDS_FORMAT" | while IFS= read -r file; do
-              echo "  google-java-format -i \"$file\""
-            done
-            exit 1
-          fi

.github/workflows/mlx.yml

@@ -47,6 +47,10 @@ jobs:
 
         ${CONDA_RUN} pip list
 
+        echo "::group::Install Python test requirements"
+        ${CONDA_RUN} pip install gguf
+        echo "::endgroup::"
+
         echo "::group::Build test runners"
         ${CONDA_RUN} cmake --build cmake-out --target op_test_runner multi_thread_test_runner -j$(( $(sysctl -n hw.ncpu) - 1 ))
         echo "::endgroup::"
@@ -76,6 +80,18 @@ jobs:
         ${CONDA_RUN} python -m executorch.backends.mlx.model_ops.test_gated_delta_rule run -v
         echo "::endgroup::"
 
+        echo "::group::Run tq_norm op tests"
+        ${CONDA_RUN} python -m executorch.backends.mlx.model_ops.test_tq_norm run -v
+        echo "::endgroup::"
+
+        echo "::group::Run tq4_compress op tests"
+        ${CONDA_RUN} python -m executorch.backends.mlx.model_ops.test_tq4_compress run -v
+        echo "::endgroup::"
+
+        echo "::group::Run tq_dequant op tests"
+        ${CONDA_RUN} python -m executorch.backends.mlx.model_ops.test_tq_dequant run -v
+        echo "::endgroup::"
+
   test-mlx-qwen35-moe:
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     with: