Merge branch 'main' into change-1223321

Author: AdrianLundell

.ci/scripts/export_model_artifact.sh

@@ -184,9 +184,17 @@ case "$HF_MODEL" in
     PREPROCESSOR_FEATURE_SIZE=""
     PREPROCESSOR_OUTPUT=""
     ;;
+  SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4)
+    MODEL_NAME="qwen3_5_moe"
+    TASK=""
+    MAX_SEQ_LEN=""
+    EXTRA_PIP=""
+    PREPROCESSOR_FEATURE_SIZE=""
+    PREPROCESSOR_OUTPUT=""
+    ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer, SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4"
     exit 1
     ;;
 esac
@@ -350,7 +358,7 @@ if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   STREAMING_ARG=""
   PREPROCESSOR_ARGS="--feature_size 128 --output_file ${OUTPUT_DIR}/preprocessor.pte"
   if [ "$USE_STREAMING" = "true" ]; then
-    STREAMING_ARG="--streaming"
+    STREAMING_ARG="--streaming --sliding-window 2048"
     PREPROCESSOR_ARGS="$PREPROCESSOR_ARGS --streaming"
   else
     PREPROCESSOR_ARGS="$PREPROCESSOR_ARGS --stack_output --max_audio_len 300"
@@ -380,6 +388,46 @@ if [ "$MODEL_NAME" = "voxtral_realtime" ]; then
   exit 0
 fi
 
+# Qwen 3.5 MoE uses a prequantized checkpoint and custom export script
+if [ "$MODEL_NAME" = "qwen3_5_moe" ]; then
+  pip install safetensors huggingface_hub
+  pip install -r examples/models/qwen3_5_moe/requirements.txt
+
+  # Download prequantized model outside OUTPUT_DIR to avoid uploading on failure
+  LOCAL_MODEL_DIR=$(mktemp -d)
+  INDUCTOR_CACHE=$(mktemp -d)
+  trap 'rm -rf "$LOCAL_MODEL_DIR" "$INDUCTOR_CACHE"' EXIT
+
+  python -c "from huggingface_hub import snapshot_download; snapshot_download('${HF_MODEL}', local_dir='${LOCAL_MODEL_DIR}')"
+
+  # Sanity check: run inference on the prequantized model
+  echo "::group::Inference sanity check"
+  python -m executorch.examples.models.qwen3_5_moe.inference \
+      --prequantized "$LOCAL_MODEL_DIR" \
+      --prompt "What is the capital of France?" \
+      --max-new-tokens 32 \
+      --temperature 0 \
+      --no-compile
+  echo "::endgroup::"
+
+  # Copy tokenizer for the runner
+  cp "$LOCAL_MODEL_DIR/tokenizer.json" "${OUTPUT_DIR}/tokenizer.json"
+
+  # Export to .pte/.ptd (short cache dir avoids objcopy symbol length issues)
+  echo "::group::Export"
+  TORCHINDUCTOR_CACHE_DIR="$INDUCTOR_CACHE" \
+  python -m executorch.examples.models.qwen3_5_moe.export \
+      --prequantized "$LOCAL_MODEL_DIR" \
+      --output-dir "${OUTPUT_DIR}"
+  echo "::endgroup::"
+
+  test -f "${OUTPUT_DIR}/model.pte"
+  test -f "${OUTPUT_DIR}/aoti_cuda_blob.ptd"
+  ls -al "${OUTPUT_DIR}"
+
+  exit 0
+fi
+
 MAX_SEQ_LEN_ARG=""
 if [ -n "$MAX_SEQ_LEN" ]; then
   MAX_SEQ_LEN_ARG="--max_seq_len $MAX_SEQ_LEN"

.ci/scripts/setup-openvino.sh

@@ -10,17 +10,41 @@ set -ex
 # shellcheck source=/dev/null
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
+# Parse arguments
+USE_NIGHTLY=false
+for arg in "$@"; do
+  case $arg in
+    --nightly) USE_NIGHTLY=true ;;
+  esac
+done
+
 # Download and install OpenVINO from release packages
-OPENVINO_VERSION="2025.3"
-OPENVINO_BUILD="2025.3.0.19807.44526285f24"
-OPENVINO_URL="https://storage.openvinotoolkit.org/repositories/openvino/packages/${OPENVINO_VERSION}/linux/openvino_toolkit_ubuntu22_${OPENVINO_BUILD}_x86_64.tgz"
+OPENVINO_VERSION="2026.0"
+OPENVINO_BUILD="2026.0.0.20965.c6d6a13a886"
+OPENVINO_STABLE_URL="https://storage.openvinotoolkit.org/repositories/openvino/packages/${OPENVINO_VERSION}/linux/openvino_toolkit_ubuntu22_${OPENVINO_BUILD}_x86_64.tgz"
+
+OPENVINO_NIGHTLY_BUILD_ID="2026.1.0-21310-c694fbc2b6d"
+OPENVINO_NIGHTLY_BUILD="2026.1.0.dev20260312"
+OPENVINO_NIGHTLY_URL="https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/${OPENVINO_NIGHTLY_BUILD_ID}/openvino_toolkit_ubuntu22_${OPENVINO_NIGHTLY_BUILD}_x86_64.tgz"
+
+if [ "${USE_NIGHTLY}" = true ]; then
+  OPENVINO_URL="${OPENVINO_NIGHTLY_URL}"
+  OPENVINO_EXTRACTED_DIR="openvino_toolkit_ubuntu22_${OPENVINO_NIGHTLY_BUILD}_x86_64"
+  echo "Using OpenVINO nightly build: ${OPENVINO_NIGHTLY_BUILD_ID}"
+else
+  OPENVINO_URL="${OPENVINO_STABLE_URL}"
+  OPENVINO_EXTRACTED_DIR="openvino_toolkit_ubuntu22_${OPENVINO_BUILD}_x86_64"
+  echo "Using OpenVINO stable release: ${OPENVINO_BUILD}"
+fi
 
 curl -Lo /tmp/openvino_toolkit.tgz --retry 3 --fail ${OPENVINO_URL}
 tar -xzf /tmp/openvino_toolkit.tgz
-mv openvino_toolkit_ubuntu22_${OPENVINO_BUILD}_x86_64 openvino
+mv "${OPENVINO_EXTRACTED_DIR}" openvino
 
+set +u
 source openvino/setupvars.sh
-cd backends/openvino
-pip install -r requirements.txt
-cd scripts
+set -u
+pip install -r backends/openvino/requirements.txt
+pushd backends/openvino/scripts
 ./openvino_build.sh --enable_python
+popd

.ci/scripts/test_backend.sh

@@ -78,6 +78,12 @@ if [[ "$FLOW" == *arm* ]]; then
     fi
 fi
 
+if [[ "$FLOW" == *openvino* ]]; then
+    # Setup OpenVINO environment
+    source .ci/scripts/setup-openvino.sh --nightly
+    EXTRA_BUILD_ARGS+=" -DEXECUTORCH_BUILD_OPENVINO=ON"
+fi
+
 if [[ $IS_MACOS -eq 1 ]]; then
     SETUP_SCRIPT=.ci/scripts/setup-macos.sh
 else

.ci/scripts/test_model_e2e.sh

@@ -216,9 +216,21 @@ case "$HF_MODEL" in
     AUDIO_FILE="test_audio.wav"
     IMAGE_PATH=""
     ;;
+  SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4)
+    MODEL_NAME="qwen3_5_moe"
+    RUNNER_TARGET="qwen3_5_moe_runner"
+    RUNNER_PATH="qwen3_5_moe"
+    EXPECTED_OUTPUT="Paris"
+    PREPROCESSOR=""
+    TOKENIZER_URL=""
+    TOKENIZER_FILE="tokenizer.json"
+    AUDIO_URL=""
+    AUDIO_FILE=""
+    IMAGE_PATH=""
+    ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, nvidia/diar_streaming_sortformer_4spk-v2, openai/whisper series (whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}), google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, nvidia/diar_streaming_sortformer_4spk-v2, openai/whisper series (whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}), google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer, SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4"
     exit 1
     ;;
 esac
@@ -232,7 +244,7 @@ echo "::group::Prepare $MODEL_NAME Artifacts"
 
 
 # Download tokenizer files (skip for models that bundle tokenizer in export or do not use one)
-if [ "$MODEL_NAME" != "parakeet" ] && [ "$MODEL_NAME" != "voxtral_realtime" ] && [ "$MODEL_NAME" != "sortformer" ] && [ "$MODEL_NAME" != "dinov2" ]; then
+if [ "$MODEL_NAME" != "parakeet" ] && [ "$MODEL_NAME" != "voxtral_realtime" ] && [ "$MODEL_NAME" != "sortformer" ] && [ "$MODEL_NAME" != "dinov2" ] && [ "$MODEL_NAME" != "qwen3_5_moe" ]; then
   if [ "$TOKENIZER_FILE" != "" ]; then
     curl -L $TOKENIZER_URL/$TOKENIZER_FILE -o $MODEL_DIR/$TOKENIZER_FILE
   else
@@ -341,6 +353,9 @@ EOF
       RUNNER_ARGS="$RUNNER_ARGS --data_path ${MODEL_DIR}/aoti_cuda_blob.ptd"
     fi
     ;;
+  qwen3_5_moe)
+    RUNNER_ARGS="$RUNNER_ARGS --tokenizer_path ${MODEL_DIR}/$TOKENIZER_FILE --prompt 'What is the capital of France?' --max_new_tokens 128 --temperature 0"
+    ;;
   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"
     # Add CUDA data path if present
@@ -359,7 +374,7 @@ EOF
     ;;
 esac
 
-OUTPUT=$($RUNNER_BIN $RUNNER_ARGS 2>&1)
+OUTPUT=$(eval $RUNNER_BIN $RUNNER_ARGS 2>&1)
 EXIT_CODE=$?
 set -e
 

.ci/scripts/test_model_e2e_windows.ps1

@@ -135,7 +135,19 @@ try {
     Write-Host "::group::Check CUDA toolchain"
     $nvccOutput = nvcc --version | Out-String
     Write-Host $nvccOutput
-    nvidia-smi
+    $nvidiaSmiCmd = Get-Command nvidia-smi -ErrorAction SilentlyContinue
+    if ($null -eq $nvidiaSmiCmd) {
+        Write-Host "nvidia-smi not available (command not found; driver may not be installed)"
+    }
+    else {
+        try {
+            nvidia-smi
+        }
+        catch {
+            Write-Host "nvidia-smi failed (driver or GPU issue). Error details:"
+            Write-Host $_
+        }
+    }
     if (-not [string]::IsNullOrWhiteSpace($ExpectedCudaVersion)) {
         $versionMatch = [Regex]::Match($nvccOutput, "release\s+(\d+\.\d+)")
         if (-not $versionMatch.Success) {

.ci/scripts/wheel/test_linux.py

@@ -11,18 +11,25 @@
 from examples.models import Backend, Model
 
 if __name__ == "__main__":
-    # On Linux x86_64 the wheel is built with the Qualcomm backend.
-    # Verify that it was registered correctly.
-    if platform.system() == "Linux" and platform.machine() in ("x86_64", "amd64"):
+    if platform.system() == "Linux":
         from executorch.extension.pybindings.portable_lib import (
             _get_registered_backend_names,
         )
 
         registered = _get_registered_backend_names()
+
+        # QNN backend is only available on x86_64.
+        if platform.machine() in ("x86_64", "amd64"):
+            assert (
+                "QnnBackend" in registered
+            ), f"QnnBackend not found in registered backends: {registered}"
+            print("✓ QnnBackend is registered")
+
+        # OpenVINO backend is available on all Linux architectures.
         assert (
-            "QnnBackend" in registered
-        ), f"QnnBackend not found in registered backends: {registered}"
-        print("✓ QnnBackend is registered")
+            "OpenvinoBackend" in registered
+        ), f"OpenvinoBackend not found in registered backends: {registered}"
+        print("✓ OpenvinoBackend is registered")
 
     test_base.run_tests(
         model_tests=[

.ci/scripts/wheel/test_linux_aarch64.py

@@ -12,6 +12,20 @@
     # 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
+
+    from executorch.extension.pybindings.portable_lib import (
+        _get_registered_backend_names,
+    )
+
+    registered = _get_registered_backend_names()
+
+    # OpenVINO backend uses dlopen (no build-time SDK dependency), so it
+    # is compiled into the wheel on all Linux architectures.
+    assert (
+        "OpenvinoBackend" in registered
+    ), f"OpenvinoBackend not found in registered backends: {registered}"
+    print("✓ OpenvinoBackend is registered")
+
     test_base.run_tests(
         model_tests=[
             test_base.ModelTest(

.github/workflows/_test_backend.yml

@@ -4,7 +4,7 @@ on:
   workflow_call:
     inputs:
       backend:
-        description: 'Backend to test (xnnpack, coreml, vulkan, qnn)'
+        description: 'Backend to test (xnnpack, coreml, vulkan, qnn, openvino)'
         required: true
         type: string
       flows:

.github/workflows/_test_cadence.yml

@@ -0,0 +1,77 @@
+name: Test Cadence
+
+permissions:
+  id-token: write
+  contents: read
+
+on:
+  workflow_call:
+    inputs:
+      docker-image:
+        description: 'Docker image to use'
+        required: false
+        type: string
+        default: ci-image:executorch-ubuntu-22.04-clang12
+      runner:
+        description: 'Runner type'
+        required: false
+        type: string
+        default: linux.8xlarge.memory
+      ref:
+        description: 'Git ref to checkout'
+        required: false
+        type: string
+        default: ${{ github.sha }}
+      timeout:
+        description: 'Job timeout in minutes'
+        required: false
+        type: number
+        default: 90
+
+jobs:
+  test-aot:
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    with:
+      job-name: test-aot
+      runner: ${{ inputs.runner }}
+      docker-image: ${{ inputs.docker-image }}
+      submodules: recursive
+      ref: ${{ inputs.ref }}
+      timeout: ${{ inputs.timeout }}
+      script: |
+        set -eux
+        conda create -y -n cadence_test python=3.12 > /dev/null
+        conda activate cadence_test
+
+        ./install_requirements.sh > /dev/null
+        pip install -e . --no-build-isolation > /dev/null
+        pip install beartype later pyre_extensions pytest-xdist
+
+        python -m pytest backends/cadence/aot/tests/ -v -n auto
+
+  test-ops:
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    with:
+      job-name: test-ops
+      runner: ${{ inputs.runner }}
+      docker-image: ${{ inputs.docker-image }}
+      submodules: recursive
+      ref: ${{ inputs.ref }}
+      timeout: ${{ inputs.timeout }}
+      download-artifact: cadence-runner-build
+      script: |
+        set -eux
+        conda create -y -n cadence_test python=3.12 > /dev/null
+        conda activate cadence_test
+
+        ./install_requirements.sh > /dev/null
+        pip install -e . --no-build-isolation > /dev/null
+        pip install beartype later pyre_extensions pytest-xdist
+
+        # Use the pre-built runner from the build job
+        mkdir -p cmake-out/backends/cadence
+        cp "${RUNNER_ARTIFACT_DIR}/cadence_runner" cmake-out/backends/cadence/cadence_runner
+        chmod +x cmake-out/backends/cadence/cadence_runner
+
+        export PYTHONPATH="${PYTHONPATH:-}:$(pwd)/backends/cadence/utils/FACTO"
+        python -m pytest examples/cadence/operators/ -v -n auto

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

@@ -145,9 +145,9 @@ jobs:
         export BUILD_AAR_DIR=aar-out
         bash scripts/build_android_library.sh
         mkdir -p "${ARTIFACTS_DIR_NAME}"
-        cp aar-out/executorch.aar "${ARTIFACTS_DIR_NAME}/executorch.aar"
+        cp aar-out/executorch.aar "${ARTIFACTS_DIR_NAME}/executorch-${FLAVOR}.aar"
 
-        shasum -a 256 "${ARTIFACTS_DIR_NAME}/executorch.aar"
+        shasum -a 256 "${ARTIFACTS_DIR_NAME}/executorch-${FLAVOR}.aar"
 
         # Publish to maven staging
         UPLOAD_TO_MAVEN="${{ inputs.upload_to_maven }}"
@@ -165,18 +165,13 @@ jobs:
       contents: read
     steps:
       - name: configure aws credentials
-        uses: aws-actions/configure-aws-credentials@v1.7.0
+        uses: aws-actions/configure-aws-credentials@v4
         with:
           role-to-assume: arn:aws:iam::308535385114:role/gha_executorch_upload-frameworks-android
           aws-region: us-east-1
       - name: Upload AAR RC to AWS S3
         shell: bash
         run: |
-          wget https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/executorch.aar
-          shasum -a 256 executorch.aar > executorch.aar.sha256sums
-
-          pip install awscli==1.32.18
-          AWS_CMD="aws s3 cp"
           VERSION="${{ inputs.version }}"
           FLAVOR="${{ inputs.flavor }}"
           if [ -z "$VERSION" ]; then
@@ -185,5 +180,11 @@ jobs:
           if [ -z "$FLAVOR" ]; then
             FLAVOR="xnnpack"
           fi
+          wget https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/executorch-${FLAVOR}.aar
+          mv executorch-${FLAVOR}.aar executorch.aar
+          shasum -a 256 executorch.aar > executorch.aar.sha256sums
+
+          pip install awscli==1.32.18
+          AWS_CMD="aws s3 cp"
           ${AWS_CMD} executorch.aar s3://ossci-android/executorch/release/${VERSION}-${FLAVOR}/executorch.aar --acl public-read
           ${AWS_CMD} executorch.aar.sha256sums s3://ossci-android/executorch/release/${VERSION}-${FLAVOR}/executorch.aar.sha256sums --acl public-read