Merge branch 'main' into mlx-delegate-part3

Author: metascroy

.ci/scripts/export_model_artifact.sh

@@ -358,7 +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"
@@ -424,6 +424,7 @@ if [ "$MODEL_NAME" = "qwen3_5_moe" ]; then
   test -f "${OUTPUT_DIR}/model.pte"
   test -f "${OUTPUT_DIR}/aoti_cuda_blob.ptd"
   ls -al "${OUTPUT_DIR}"
+
   exit 0
 fi
 

.ci/scripts/setup-qnn-deps.sh

@@ -11,4 +11,5 @@ source "$(dirname "${BASH_SOURCE[0]}")/../../backends/qualcomm/scripts/install_q
 
 setup_libcpp 12
 setup_android_ndk
-install_qnn
+install_qnn
+pip install -r backends/qualcomm/requirements.txt

.ci/scripts/test_model_e2e.sh

@@ -354,7 +354,7 @@ EOF
     fi
     ;;
   qwen3_5_moe)
-    RUNNER_ARGS="$RUNNER_ARGS --tokenizer_path ${MODEL_DIR}/$TOKENIZER_FILE --prompt 'What is the capital of France?' --max_new_tokens 32"
+    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"

.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/test_wheel_package_qnn.sh

@@ -177,6 +177,9 @@ PY
   "$PIPBIN" install . --no-build-isolation
   popd > /dev/null
 
+  # Install qualcomm backend dependencies  
+  "$PIPBIN" install -r "$REPO_ROOT/backends/qualcomm/requirements.txt"
+
   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')"

.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(

.githooks/README.md

@@ -1,57 +1,63 @@
 # Git Hooks
 
-This directory contains Git hooks for the ExecuTorch repository.
+This directory contains Git hooks for the ExecuTorch repository. It is used as
+`core.hooksPath`, so git looks here instead of `.git/hooks/`.
 
-## Pre-commit Hook
+## Hooks
 
-The pre-commit hook automatically updates the PyTorch commit pin in `.ci/docker/ci_commit_pins/pytorch.txt` whenever `torch_pin.py` is modified.
+### pre-commit
 
-### How It Works
+Runs on every commit:
 
-1. When you commit changes to `torch_pin.py`, the hook detects the change
-2. It parses the `NIGHTLY_VERSION` field (e.g., `dev20251004`)
-3. Converts it to a date string (e.g., `2025-10-04`)
-4. Fetches the corresponding commit hash from the PyTorch nightly branch at https://github.com/pytorch/pytorch/tree/nightly
-5. Updates `.ci/docker/ci_commit_pins/pytorch.txt` with the new commit hash
-6. Automatically stages the updated file for commit
+1. **torch_pin sync** — when `torch_pin.py` is staged, updates the PyTorch commit
+   pin in `.ci/docker/ci_commit_pins/pytorch.txt` and syncs grafted c10 files.
+2. **lintrunner** — runs `lintrunner -a --revision HEAD^ --skip MYPY` on changed
+   files. Auto-fixes formatting and blocks on lint errors. Soft-fails if lintrunner
+   is not installed. Runs `lintrunner init` automatically when `.lintrunner.toml`
+   changes.
 
-### Installation
+### pre-push
 
-To install the Git hooks, run:
+Delegates to `.git/hooks/pre-push` if one exists. This allows backend-specific
+pre-push hooks (e.g., ARM's license and commit message checks) to work alongside
+the repo-wide hooks.
+
+## Installation
+
+Hooks are installed automatically by `./install_executorch.sh`.
+
+To install manually:
 
 ```bash
-.githooks/install.sh
+git config core.hooksPath .githooks
 ```
 
-This will copy the pre-commit hook to `.git/hooks/` and make it executable.
+### ARM backend pre-push
 
-### Manual Usage
-
-You can also run the update script manually at any time:
+ARM contributors should additionally install the ARM-specific pre-push hook:
 
 ```bash
-python .github/scripts/update_pytorch_pin.py
+cp backends/arm/scripts/pre-push .git/hooks/
 ```
 
-### Uninstalling
+## Bypassing
 
-To remove the pre-commit hook:
+To skip hooks for a single commit or push:
 
 ```bash
-rm .git/hooks/pre-commit
+git commit --no-verify
+git push --no-verify
 ```
 
 ## Troubleshooting
 
-If the hook fails during a commit:
+If the torch_pin hook fails:
 
 1. Check that Python 3 is available in your PATH
 2. Ensure you have internet connectivity to fetch commits from GitHub
 3. Verify that the `NIGHTLY_VERSION` in `torch_pin.py` is in the correct format (`devYYYYMMDD`)
-4. Make sure the corresponding nightly release exists in the PyTorch nightly branch
 
-You can run the script manually to see detailed error messages:
+If lintrunner fails:
 
-```bash
-python .github/scripts/update_pytorch_pin.py
-```
+1. Run `lintrunner init` to install linter tools
+2. Check that your virtual environment is activated

.githooks/install.sh

@@ -27,4 +27,68 @@ if git diff --cached --name-only | grep -q "^torch_pin.py$"; then
     fi
 fi
 
-exit 0
+# --- lintrunner ---
+
+if ! command -v lintrunner >/dev/null 2>&1; then
+    echo "Warning: lintrunner not found. Skipping lint checks."
+    echo "Install with: pip install lintrunner lintrunner-adapters && lintrunner init"
+    exit 0
+fi
+
+if [ ! -f .lintrunner.toml ]; then
+    echo "Warning: .lintrunner.toml not found. Skipping lint checks."
+    exit 0
+fi
+
+git_dir=$(git rev-parse --git-dir)
+
+# Portable hash: sha256sum (Linux) or shasum (macOS)
+if command -v sha256sum >/dev/null 2>&1; then
+    toml_hash=$(sha256sum .lintrunner.toml | cut -d' ' -f1)
+else
+    toml_hash=$(shasum -a 256 .lintrunner.toml | cut -d' ' -f1)
+fi
+stored_hash=""
+[ -f "${git_dir}/.lintrunner_init_hash" ] && stored_hash=$(cat "${git_dir}/.lintrunner_init_hash")
+
+if [ "${toml_hash}" != "${stored_hash}" ]; then
+    echo "Running lintrunner init..."
+    if lintrunner init; then
+        echo "${toml_hash}" > "${git_dir}/.lintrunner_init_hash"
+    else
+        echo "lintrunner init failed. Run 'lintrunner init' manually."
+        exit 1
+    fi
+fi
+
+staged_files=$(git diff --cached --name-only --diff-filter=ACMR)
+
+# Use HEAD^ if it exists (skip on initial commit)
+revision_flag="--revision HEAD^"
+if ! git rev-parse HEAD^ >/dev/null 2>&1; then
+    revision_flag=""
+fi
+
+lintrunner -a $revision_flag --skip MYPY
+lint_status=$?
+
+# Check if lintrunner modified any staged files. If so, block the commit
+# so the user can review the changes before committing.
+files_modified=0
+while IFS= read -r path; do
+    [ -z "${path}" ] && continue
+    if ! git diff --quiet -- "${path}" 2>/dev/null; then
+        files_modified=1
+        break
+    fi
+done <<< "${staged_files}"
+
+if [ $files_modified -eq 1 ]; then
+    echo "Lintrunner modified files. Review with 'git diff', then 'git add -u && git commit'."
+    exit 1
+fi
+
+if [ $lint_status -ne 0 ]; then
+    echo "Lint errors found. Fix them and try again, or use 'git commit --no-verify' to skip."
+    exit 1
+fi