Update

[ghstack-poisoned]

Author: JCNTH

.ci/scripts/test-cadence-xtensa.sh

@@ -0,0 +1,84 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+#
+# Build the Cadence Xtensa op-level gtest tests for the configured backend and
+# run them on the Instruction Set Simulator (xt-run).
+#
+# Requires the Xtensa toolchain env to already be set (run
+# .ci/scripts/setup-xtensa-tools.sh <backend> first): XTENSA_TOOLCHAIN,
+# TOOLCHAIN_VER, XTENSA_CORE, CADENCE_OPT_FLAG, and xt-clang/xt-run on PATH.
+#
+# Unlike build-cadence-xtensa.sh (the runner, built -fno-exceptions -fno-rtti),
+# the gtest tests need exceptions + RTTI, so those flags are NOT set here.
+
+set -euo pipefail
+
+: "${XTENSA_TOOLCHAIN:?run setup-xtensa-tools.sh first}"
+: "${TOOLCHAIN_VER:?run setup-xtensa-tools.sh first}"
+: "${XTENSA_CORE:?run setup-xtensa-tools.sh first}"
+: "${CADENCE_OPT_FLAG:?run setup-xtensa-tools.sh first}"
+
+# Map the optimized-kernel flag to the backend dir + gtest target name.
+case "${CADENCE_OPT_FLAG}" in
+  EXECUTORCH_NNLIB_OPT) TARGET_DIR=hifi ;;
+  EXECUTORCH_VISION_OPT) TARGET_DIR=vision ;;
+  EXECUTORCH_FUSION_G3_OPT) TARGET_DIR=fusion_g3 ;;
+  *)
+    echo "ERROR: unknown CADENCE_OPT_FLAG='${CADENCE_OPT_FLAG}'" >&2
+    exit 1
+    ;;
+esac
+TEST_TARGET="cadence_${TARGET_DIR}_op_tests"
+TEST_ELF="cmake-out/backends/cadence/${TARGET_DIR}/operators/tests/${TEST_TARGET}"
+
+NPROC=$(nproc)
+echo "=== building ${TEST_TARGET} for ${XTENSA_CORE} (${CADENCE_OPT_FLAG}) ==="
+xt-clang --version | head -1
+
+rm -rf cmake-out
+cmake \
+  -DCMAKE_TOOLCHAIN_FILE=./backends/cadence/cadence.cmake \
+  -DCMAKE_INSTALL_PREFIX=cmake-out \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DEXECUTORCH_BUILD_CADENCE=ON \
+  "-D${CADENCE_OPT_FLAG}=ON" \
+  -DEXECUTORCH_BUILD_PORTABLE_OPS=ON \
+  -DEXECUTORCH_BUILD_CADENCE_OP_TESTS=ON \
+  -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=OFF \
+  -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
+  -DEXECUTORCH_ENABLE_LOGGING=ON \
+  -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
+  -DEXECUTORCH_BUILD_CPUINFO=OFF \
+  -DEXECUTORCH_USE_DL=OFF \
+  -DEXECUTORCH_BUILD_KERNELS_LLM=OFF \
+  -DEXECUTORCH_BUILD_DEVTOOLS=OFF \
+  -DHAVE_FNMATCH_H=OFF \
+  -DFLATCC_ALLOW_WERROR=OFF \
+  -DPYTHON_EXECUTABLE="$(which python3)" \
+  -Bcmake-out .
+
+cmake --build cmake-out --target "${TEST_TARGET}" -j"${NPROC}"
+
+if [[ ! -f "${TEST_ELF}" ]]; then
+  echo "ERROR: ${TEST_ELF} was not produced" >&2
+  exit 1
+fi
+
+echo "=== running ${TEST_TARGET} on xt-run ==="
+LOG=$(mktemp)
+# --exit_with_target_code propagates gtest_main's exit code, so a failing test
+# fails this step; also assert on the gtest summary lines as a backstop.
+xt-run --turbo --exit_with_target_code "${TEST_ELF}" 2>&1 | tee "${LOG}"
+if grep -q "\[  FAILED  \]" "${LOG}"; then
+  echo "ERROR: gtest reported failures for ${TEST_TARGET}" >&2
+  exit 1
+fi
+if ! grep -q "\[  PASSED  \]" "${LOG}"; then
+  echo "ERROR: ${TEST_TARGET} did not report a gtest PASSED summary" >&2
+  exit 1
+fi
+echo "Cadence ${TARGET_DIR} op tests passed on ${XTENSA_CORE}."

.ci/scripts/wheel/pre_build_script.sh

@@ -122,4 +122,8 @@ if [[ "${EXECUTORCH_BUILD_VULKAN:-0}" != "0" \
     echo "${_glslc_prefix}/bin" >> "${GITHUB_PATH}"
     echo "glslc installed: $(command -v glslc)"
   fi
+else
+  # This is the Vulkan equivalent of the Windows CUDA force-off above (#20527).
+  export CMAKE_ARGS="${CMAKE_ARGS:-} -DEXECUTORCH_BUILD_VULKAN=OFF"
+  echo "CMAKE_ARGS=${CMAKE_ARGS}" >> "${GITHUB_ENV}"
 fi

.ci/scripts/wheel/test_base.py

@@ -1,5 +1,6 @@
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -40,6 +41,23 @@ class ModelTest:
     backend: Backend
 
 
+def test_cmsis_nn_install():
+    import executorch.backends.cortex_m.library.cmsis_nn as cmsis_nn
+
+    buf_size = cmsis_nn.convolve_wrapper_buffer_size(
+        cmsis_nn.Backend.MVE,
+        cmsis_nn.DataType.A8W8,
+        input_nhwc=[1, 8, 8, 16],
+        filter_nhwc=[8, 3, 3, 16],
+        output_nhwc=[1, 6, 6, 8],
+        padding_hw=[0, 0],
+        stride_hw=[1, 1],
+        dilation_hw=[1, 1],
+    )
+
+    assert buf_size == 576
+
+
 def run_tests(model_tests: List[ModelTest]) -> None:
     # Test that we can import the portable_lib module - verifies RPATH is correct
     print("Testing portable_lib import...")

.ci/scripts/wheel/test_linux.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -38,6 +39,8 @@
         else:
             print("⚠ VulkanBackend not registered (expected for the default wheel)")
 
+        test_base.test_cmsis_nn_install()
+
     test_base.run_tests(
         model_tests=[
             test_base.ModelTest(

.ci/scripts/wheel/test_macos.py

@@ -1,6 +1,7 @@
 #!/usr/bin/env python
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -9,6 +10,8 @@
 from examples.models import Backend, Model
 
 if __name__ == "__main__":
+    test_base.test_cmsis_nn_install()
+
     test_base.run_tests(
         model_tests=[
             test_base.ModelTest(

.ci/scripts/wheel/test_windows.py

@@ -1,13 +1,16 @@
 #!/usr/bin/env python
 # Copyright (c) Meta Platforms, Inc. and affiliates.
 # All rights reserved.
+# Copyright 2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
 import platform
 from typing import List
 
+import test_base
+
 import torch
 from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
 from executorch.examples.models import Backend, Model, MODEL_NAME_TO_MODEL
@@ -74,6 +77,8 @@ def run_tests(model_tests: List[ModelTest]) -> None:
         else:
             print("⚠ VulkanBackend not registered (expected for the default wheel)")
 
+        test_base.test_cmsis_nn_install()
+
     run_tests(
         model_tests=[
             ModelTest(

.claude/skills/qualcomm/new_op_development.md

@@ -66,7 +66,7 @@ class MyOpVisitor(NodeVisitor):
 - Static params: `weight = get_parameter(self.get_node(node.args[1]), self.edge_program)`
 - Scalar params → `AddScalarParam`; Array params → `AddTensorParam`
 - Data types: axis/dims=`UINT_32`, epsilon=`FLOAT_32`, booleans=`BOOL_8`
-- Int64 index tensors: use `.to(torch.int32)` in builder + add op to `I64_IN_OPS` in `_passes/i64_to_i32.py` for CPU fallback safety (see `op_gather.py` pattern)
+- Int64 index tensors: If the op **requires** int64 for PyTorch tracing validation (like `gather`, `scatter`), add to `I64_IN_OPS` in `i64_to_i32.py` + `.to(torch.int32)` in builder. If the op **accepts** int32 (like `index_select`), produce int32 directly via `dtype=torch.int32` — no `I64_IN_OPS` entry needed.
 
 ## Step 5: Register Builder (`builders/__init__.py`)
 
@@ -237,7 +237,7 @@ class DecomposeMyOp(ExportPass):
 - **Multi-output ops**: Use `wrapper_idx=i` + `getitem` skip op
 - **Negative dims**: QNN needs positive → `dim = dim % len(shape)`
 - **QCOM_AXIS_ORDER**: `LayoutTransform` permutes NCHW→NHWC; remap axis with `.index(dim)`. `get_tensor()` auto-permutes data.
-- **Int64 indices**: Add to `I64_IN_OPS` in `i64_to_i32.py` + `.to(torch.int32)` in builder (see `op_gather.py`)
+- **Int64 indices**: Only add to `I64_IN_OPS` if the op **requires** int64 at tracing time (e.g., `gather`, `scatter`). If the op accepts int32 (e.g., `index_select`), produce int32 directly in the decomposition pass. Check PyTorch docs for actual dtype requirements.
 - **Recompose passes**: Detect primitive sequences and replace with single native op. Ref: `recompose_pixel_unshuffle.py`
 - **`partition/common_defs.py`**: Remove op from `to_be_implemented_operator` when adding support
 - **HTP doc bugs**: If runtime fails but docs say supported → test on-device always.

.github/workflows/_xtensa_test.yml

@@ -0,0 +1,84 @@
+# Reusable: build + run the Cadence Xtensa op-level gtest tests for one core on
+# the Instruction Set Simulator (xt-run). Mirrors _xtensa_build.yml's native
+# OIDC + docker-run skeleton (running xt-run needs the same licensed toolchain),
+# then builds the gtest op-test ELF and runs it on the simulator. The runner
+# cross-compile and these op tests are separate build configs (the tests need
+# exceptions/RTTI that the runner build disables), so this is a self-contained
+# job rather than consuming the runner artifact.
+name: xtensa-test
+
+on:
+  workflow_call:
+    inputs:
+      backend:
+        description: "Cadence backend to test (hifi4 | vision | fusion_g3)"
+        required: true
+        type: string
+      ref:
+        description: "Git ref to check out"
+        required: false
+        type: string
+        default: ""
+
+jobs:
+  test:
+    name: ${{ inputs.backend }}
+    runs-on: linux.2xlarge
+    environment: cadence
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - name: Checkout executorch
+        uses: actions/checkout@v4
+        with:
+          submodules: recursive
+          ref: ${{ inputs.ref }}
+
+      - name: Calculate docker image
+        id: calculate-docker-image
+        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
+        with:
+          docker-image-name: ci-image:executorch-ubuntu-22.04-clang12
+
+      - name: Pull docker image
+        run: docker pull "${{ steps.calculate-docker-image.outputs.docker-image }}"
+
+      - name: Assume Cadence artifacts role (host OIDC)
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          role-to-assume: ${{ vars.CADENCE_CI_AWS_ROLE }}
+          aws-region: ${{ vars.CADENCE_CI_AWS_REGION }}
+
+      - name: Build and run op tests on xt-run
+        env:
+          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
+          BACKEND: ${{ inputs.backend }}
+          XTENSA_S3_BUCKET: ${{ vars.CADENCE_CI_S3_BUCKET }}
+        shell: bash
+        run: |
+          set -eux
+          # OIDC/role assumption already happened on the host above; pass the
+          # resulting AWS creds and the store/backend into the CI image, where
+          # the toolchain download + op-test build + xt-run happen.
+          docker run --rm \
+            -e BACKEND -e XTENSA_S3_BUCKET \
+            -e AWS_ACCESS_KEY_ID -e AWS_SECRET_ACCESS_KEY -e AWS_SESSION_TOKEN \
+            -e AWS_DEFAULT_REGION -e AWS_REGION \
+            -v "${GITHUB_WORKSPACE}:/work/executorch" -w /work/executorch \
+            "${DOCKER_IMAGE}" \
+            bash -c '
+              set -exo pipefail
+              eval "$(/opt/conda/bin/conda shell.bash hook)"
+              conda activate "$(conda env list --json | jq -r ".envs | .[-1]")"
+              ./install_requirements.sh > /dev/null
+              pip install --quiet awscli
+              # hifi4/fusion_g3 optimized kernels need the foss-xtensa nnlib
+              # sources, which are not vendored in executorch; the cadence
+              # installer clones them. vision has no nnlib dependency.
+              if [ "${BACKEND}" != "vision" ]; then
+                backends/cadence/install_requirements.sh
+              fi
+              source .ci/scripts/setup-xtensa-tools.sh "${BACKEND}"
+              .ci/scripts/test-cadence-xtensa.sh
+            '

.github/workflows/build-cadence-runner.yml

@@ -87,6 +87,25 @@ jobs:
       backend: hifi4
       ref: ${{ (github.event_name == 'pull_request' || github.event_name == 'pull_request_target') && github.event.pull_request.head.sha || github.sha }}
 
+  # Op-level gtest tests on the Xtensa ISS, mirroring cpu-build -> cpu-test. The
+  # op tests are a self-contained cross-compile (the gtests need exceptions and
+  # RTTI that the runner build disables), so this does not consume hifi-build's
+  # artifact; needs: hifi-build only to fail fast when the backend cannot build.
+  hifi-op-test:
+    needs: hifi-build
+    if: >-
+      github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository) ||
+      (github.event_name == 'pull_request_target' && github.event.pull_request.head.repo.full_name != github.repository &&
+      contains(github.event.pull_request.labels.*.name, 'CLA Signed') && contains(github.event.pull_request.labels.*.name, 'meta-exported'))
+    permissions:
+      id-token: write
+      contents: read
+    uses: ./.github/workflows/_xtensa_test.yml
+    with:
+      backend: hifi4
+      ref: ${{ (github.event_name == 'pull_request' || github.event_name == 'pull_request_target') && github.event.pull_request.head.sha || github.sha }}
+
   vision-build:
     if: >-
       github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' ||

.github/workflows/cuda.yml

@@ -213,6 +213,40 @@ jobs:
         pip install gguf
         python -m pytest examples/models/gemma4_31b/quant/tests/ examples/models/gemma4_31b/tests/ --ignore=examples/models/gemma4_31b/tests/test_mlx_pipeline.py -v -o "addopts="
 
+  unittest-cuda-runtime:
+    name: unittest-cuda-runtime
+    needs: [changed-files, run-decision]
+    if: |
+      contains(needs.changed-files.outputs.changed-files, 'backends/cuda') ||
+      contains(needs.changed-files.outputs.changed-files, 'backends/aoti') ||
+      contains(needs.changed-files.outputs.changed-files, '.github/workflows/cuda.yml') ||
+      contains(needs.changed-files.outputs.changed-files, '.ci/scripts/test-cuda-build.sh') ||
+      contains(needs.changed-files.outputs.changed-files, '.ci/scripts/export_model_artifact.sh') ||
+      contains(needs.changed-files.outputs.changed-files, '.ci/scripts/test_model_e2e.sh') ||
+      needs.run-decision.outputs.is-full-run == 'true'
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    with:
+      timeout: 90
+      runner: linux.g5.4xlarge.nvidia.gpu
+      gpu-arch-type: cuda
+      gpu-arch-version: "13.0"
+      use-custom-docker-registry: false
+      submodules: recursive
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      script: |
+        set -eux
+        bash ./install_executorch.sh
+        conda install -y -c conda-forge 'libstdcxx-ng>=12'
+        export LD_LIBRARY_PATH=/opt/conda/lib:$LD_LIBRARY_PATH
+
+        cmake --preset llm-release-cuda -DEXECUTORCH_BUILD_TESTS=ON
+        cmake --build cmake-out --target test_cuda_allocator test_cuda_mutable_state -j$(nproc)
+        ctest --test-dir cmake-out -R test_cuda_allocator --output-on-failure -V
+        ctest --test-dir cmake-out -R test_cuda_mutable_state --output-on-failure -V
+
   export-model-cuda-artifact:
     name: export-model-cuda-artifact
     # Skip this job if the pull request is from a fork (HuggingFace secrets are not available).