Merge branch 'main' into cuda-graph

Author: Gasoonjia

.ci/scripts/test_cortex_m_e2e.sh

@@ -1,5 +1,6 @@
 #!/usr/bin/env bash
 # Copyright (c) Meta Platforms, Inc. and affiliates.
+# Copyright 2026 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -18,7 +19,7 @@ mkdir -p "./cortex_m_e2e/${MODEL}"
 WORK_DIR=$(realpath "./cortex_m_e2e/${MODEL}")
 
 echo "=== Exporting ${MODEL} with cortex-m55+int8 ==="
-python -m examples.arm.aot_arm_compiler \
+python -m backends.arm.scripts.aot_arm_compiler \
     -m "${MODEL}" \
     --target=cortex-m55+int8 \
     --quantize \

.ci/scripts/test_yolo12.sh .ci/scripts/test_yolo26.sh.ci/scripts/test_yolo12.sh renamed to .ci/scripts/test_yolo26.sh

@@ -6,6 +6,7 @@
 # LICENSE file in the root directory of this source tree.
 
 set -ex
+
 # shellcheck source=/dev/null
 source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 
@@ -50,21 +51,21 @@ PT2E_QUANTIZE="${PT2E_QUANTIZE:-}"
 # Default CMake Build Type to release mode
 CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release}
 
-if [[ $# -lt 5 ]]; then # Assuming 4 mandatory args
-    echo "Expecting atleast 5 positional arguments"
-    echo "Usage: [...]"
-fi
 if [[ -z "${MODEL_NAME:-}" ]]; then
   echo "Missing model name, exiting..."
   exit 1
 fi
 
-
 if [[ -z "${MODE:-}" ]]; then
   echo "Missing mode, choose openvino or xnnpack, exiting..."
   exit 1
 fi
 
+if [[ -z "${VIDEO_PATH:-}" ]]; then
+  echo "Missing video path, exiting..."
+  exit 1
+fi
+
 if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
   PYTHON_EXECUTABLE=python3
 fi
@@ -75,21 +76,13 @@ if [[ "${MODE}" =~ .*openvino.* ]]; then
   OPENVINO=ON
   TARGET_LIBS="$TARGET_LIBS openvino_backend "
 
-  git clone https://github.com/openvinotoolkit/openvino.git
-  cd openvino && git b16b776ac119dafda51f69a80f1e6b7376d02c3b
-  git submodule update --init --recursive
-  sudo ./install_build_dependencies.sh
-  mkdir build && cd build
-  cmake .. -DCMAKE_BUILD_TYPE=Release -DENABLE_PYTHON=ON
-  make -j$(nproc)
-
-  cd ..
-  cmake --install build --prefix dist
-
-  source dist/setupvars.sh
-  cd ../backends/openvino
-  pip install -r requirements.txt
-  cd ../../
+  # Install specific OpenVINO runtime from pip.
+  $PYTHON_EXECUTABLE -m pip install --pre openvino==2026.1.0.dev20260131 --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
+  $PYTHON_EXECUTABLE -m pip install -r backends/openvino/requirements.txt
+
+  # Set OPENVINO_LIB_PATH so the C++ demo runner can also find libopenvino_c.so.
+  OPENVINO_LIB_PATH=$($PYTHON_EXECUTABLE -c "import openvino, os, glob; print(sorted(glob.glob(os.path.join(os.path.dirname(openvino.__file__), 'libs', 'libopenvino_c.so*')))[-1])")
+  export OPENVINO_LIB_PATH
 else
   OPENVINO=OFF
 fi
@@ -103,9 +96,10 @@ fi
 
 which "${PYTHON_EXECUTABLE}"
 
+TORCH_URL=https://download.pytorch.org/whl/cpu
 
-DIR="examples/models/yolo12"
-$PYTHON_EXECUTABLE -m pip install -r ${DIR}/requirements.txt
+DIR="examples/models/yolo26"
+$PYTHON_EXECUTABLE -m pip install --upgrade-strategy only-if-needed --extra-index-url "$TORCH_URL" -r ${DIR}/requirements.txt
 
 cmake_install_executorch_libraries() {
     rm -rf cmake-out
@@ -142,11 +136,11 @@ cmake_install_executorch_libraries() {
 
     echo $TARGET_LIBS
     export CMAKE_BUILD_ARGS="--target $TARGET_LIBS"
-    pip install . --no-build-isolation
+    $PYTHON_EXECUTABLE -m pip install . --no-build-isolation
 }
 
 cmake_build_demo() {
-    echo "Building yolo12 runner"
+    echo "Building yolo26 runner"
     retry cmake \
         -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
         -DUSE_OPENVINO_BACKEND="$OPENVINO" \
@@ -174,24 +168,29 @@ prepare_artifacts_upload() {
 
 
 # Export model.
-EXPORTED_MODEL_NAME="${MODEL_NAME}_fp32_${MODE}.pte"
-echo "Exporting ${EXPORTED_MODEL_NAME}"
 EXPORT_ARGS="--model_name=${MODEL_NAME} --backend=${MODE}"
+if [[ -n "${PT2E_QUANTIZE}" ]]; then
+  EXPORTED_MODEL_NAME="${MODEL_NAME}_int8_${MODE}.pte"
+  EXPORT_ARGS="${EXPORT_ARGS} --quantize --video_path=${VIDEO_PATH}"
+else
+  EXPORTED_MODEL_NAME="${MODEL_NAME}_fp32_${MODE}.pte"
+fi
+echo "Exporting ${EXPORTED_MODEL_NAME}"
 
 # Add dynamically linked library location
 cmake_install_executorch_libraries
 
-$PYTHON_EXECUTABLE -m examples.models.yolo12.export_and_validate ${EXPORT_ARGS}
+$PYTHON_EXECUTABLE -m examples.models.yolo26.export_and_validate ${EXPORT_ARGS}
 
 
 RUNTIME_ARGS="--model_path=${EXPORTED_MODEL_NAME} --input_path=${VIDEO_PATH}"
 # Check build tool.
 cmake_build_demo
-# Run yolo12 runner
+# Run yolo26 runner
 NOW=$(date +"%H:%M:%S")
-echo "Starting to run yolo12 runner at ${NOW}"
+echo "Starting to run yolo26 runner at ${NOW}"
 # shellcheck source=/dev/null
-cmake-out/examples/models/yolo12/Yolo12DetectionDemo ${RUNTIME_ARGS} > result.txt
+cmake-out/examples/models/yolo26/Yolo26DetectionDemo ${RUNTIME_ARGS} > result.txt
 NOW=$(date +"%H:%M:%S")
 echo "Finished at ${NOW}"
 

.github/workflows/pull.yml

@@ -674,6 +674,40 @@ jobs:
       build-tool: buck2
       docker-image: ci-image:executorch-ubuntu-22.04-clang12
 
+  test-qnn-buck-build-linux:
+    name: test-qnn-buck-build-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: ci-image:executorch-ubuntu-22.04-qnn-sdk
+      submodules: 'recursive'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 90
+      script: |
+        set -eux
+
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        # Download QNN SDK and get the path
+        QNN_SDK_ROOT=$(python3 backends/qualcomm/scripts/download_qnn_sdk.py --print-sdk-path)
+        echo "QNN_SDK_ROOT=${QNN_SDK_ROOT}"
+
+        # Configure Buck to find the QNN SDK
+        echo "[qualcomm]" >> .buckconfig
+        echo "  qnn_sdk_root = ${QNN_SDK_ROOT}" >> .buckconfig
+
+        # Setup buck2
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool buck2
+
+        # Build QNN backend with Buck
+        buck2 build //backends/qualcomm/...
+
   unittest-arm-backend-with-no-deps:
     name: unittest-arm-backend-with-no-deps
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
@@ -823,6 +857,8 @@ jobs:
       id-token: write
       contents: read
     strategy:
+      matrix:
+        model: [mv2, mv3, dl3]
       fail-fast: false
     with:
       runner: linux.2xlarge
@@ -834,9 +870,58 @@ jobs:
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
         conda activate "${CONDA_ENV}"
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool cmake
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_model.sh ${{ matrix.model }} "cmake" "qnn"
+
+  test-qnn-testsuite-linux:
+    name: test-qnn-testsuite-linux
+    permissions:
+      id-token: write
+      contents: read
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: qnn
+      flows: '["qnn"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true
+      runner-linux: linux.2xlarge
+
+  test-qnn-passes-linux:
+    name: test-qnn-passes-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: ci-image:executorch-ubuntu-22.04-qnn-sdk
+      submodules: 'recursive'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 30
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
+        # Source (not bash) so QNN_SDK_ROOT stays in the environment
+        PYTHON_EXECUTABLE=python source .ci/scripts/build-qnn-sdk.sh
+
+        # Editable install so the PyQnnManagerAdaptor .so built by build-qnn-sdk.sh
+        # is visible in the source tree (the _passes import chain pulls it in transitively)
+        CMAKE_ARGS="-DEXECUTORCH_BUILD_QNN=ON -DQNN_SDK_ROOT=$QNN_SDK_ROOT -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON" \
+          PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool cmake --editable true
+
+        # test_mha_to_sha imports oss_scripts/llama which transitively needs torchtune
+        pip install -r requirements-examples.txt
 
-        # placeholder for running test_qnn_delegate.py, can use matrix such that we can trigger different jobs, refers to test-llama-runner-qnn-linux
-        # reminder: make sure each job runs fast
+        # Run QNN pass unit tests
+        pytest -xvs backends/qualcomm/tests/test_passes.py
 
   test-phi-3-mini-runner-linux:
     name: test-phi-3-mini-runner-linux

.lintrunner.toml

@@ -225,6 +225,7 @@ exclude_patterns = [
     'extension/llm/tokenizers/**',
     'backends/cadence/utils/FACTO',
     'examples/cuda',
+    'examples/qualcomm',
     'kernels/portable',
     # File contains @generated
     'extension/llm/custom_ops/spinquant/fast_hadamard_transform_special.h',

README.md

@@ -204,7 +204,7 @@ ExecuTorch powers on-device AI at scale across Meta's family of apps, VR/AR devi
 
 **Multimodal:** [Llava](examples/models/llava/README.md) (vision-language), [Voxtral](examples/models/voxtral/README.md) (audio-language), [Gemma](examples/models/gemma3) (vision-language)
 
-**Vision/Speech:** [MobileNetV2](https://github.com/meta-pytorch/executorch-examples/tree/main/mv2), [DeepLabV3](https://github.com/meta-pytorch/executorch-examples/tree/main/dl3), [Whisper](examples/models/whisper/README.md) <!-- @lint-ignore -->
+**Vision/Speech:** [MobileNetV2](https://github.com/meta-pytorch/executorch-examples/tree/main/mv2), [DeepLabV3](https://github.com/meta-pytorch/executorch-examples/tree/main/dl3), [YOLO26](examples/models/yolo26/README.md), [Whisper](examples/models/whisper/README.md) <!-- @lint-ignore -->
 
 **Resources:** [`examples/`](examples/) directory • [executorch-examples](https://github.com/meta-pytorch/executorch-examples) out-of-tree demos • [Optimum-ExecuTorch](https://github.com/huggingface/optimum-executorch) for HuggingFace models • [Unsloth](https://docs.unsloth.ai/new/deploy-llms-phone) for fine-tuned LLM deployment <!-- @lint-ignore -->
 

backends/apple/metal/CMakeLists.txt

@@ -48,6 +48,7 @@ set(_aoti_metal_sources
     runtime/ops/op_linear_4bit.mm
     runtime/ops/op_mm.mm
     runtime/ops/op_sdpa.mm
+    runtime/ops/op_topk.mm
 )
 
 add_library(metal_backend STATIC ${_aoti_metal_sources})

backends/apple/metal/metal_backend.py

@@ -36,6 +36,7 @@ def get_supported_fallback_kernels(cls) -> Dict[str, Any]:
             "aoti_torch_mps_mm_out": None,
             "at::_ops::_scaled_dot_product_attention_math_for_mps::call": None,
             "torchao::_linear_fp_act_4bit_weight": None,
+            "at::_ops::topk::call": None,
         }
 
     @classmethod

backends/apple/metal/runtime/ops/op_sdpa.mm

@@ -226,7 +226,8 @@
 #define INSTANTIATE_SDPA_VECTOR_HEADS(DTYPE)        \
   INSTANTIATE_SDPA_VECTOR(DTYPE, 64, 64);           \
   INSTANTIATE_SDPA_VECTOR(DTYPE, 96, 96);           \
-  INSTANTIATE_SDPA_VECTOR(DTYPE, 128, 128);
+  INSTANTIATE_SDPA_VECTOR(DTYPE, 128, 128);         \
+  INSTANTIATE_SDPA_VECTOR(DTYPE, 256, 256);
 
 INSTANTIATE_SDPA_VECTOR_HEADS(float);
 INSTANTIATE_SDPA_VECTOR_HEADS(bfloat);
@@ -430,11 +431,11 @@ AOTITorchError aoti_torch_mps__scaled_dot_product_attention_math_for_mps(
         throw std::runtime_error("Unsupported dtype for Metal SDPA kernel");
       }
 
-      // Select head_dim - must match exactly one of the supported sizes (64, 96, 128)
+      // Select head_dim - must match exactly one of the supported sizes (64, 96, 128, 256)
       int64_t head_dim = headSize;
-      if (head_dim != 64 && head_dim != 96 && head_dim != 128) {
-        ET_LOG(Error, "aoti_torch_mps__scaled_dot_product_attention_math_for_mps: Unsupported head_dim %lld (must be 64, 96, or 128)", head_dim);
-        throw std::runtime_error("Unsupported head_dim for Metal SDPA kernel - must be exactly 64, 96, or 128");
+      if (head_dim != 64 && head_dim != 96 && head_dim != 128 && head_dim != 256) {
+        ET_LOG(Error, "aoti_torch_mps__scaled_dot_product_attention_math_for_mps: Unsupported head_dim %lld (must be 64, 96, 128, or 256)", head_dim);
+        throw std::runtime_error("Unsupported head_dim for Metal SDPA kernel - must be exactly 64, 96, 128, or 256");
       }
 
       std::string kernel_name = "sdpa_vector_" + type_name + "_" + std::to_string(head_dim) + "_" + std::to_string(head_dim);