Merge branch 'main' of https://github.com/pytorch/executorch into export-D82171193

Author: AdrianLundell

.ci/scripts/test_torchao_huggingface_checkpoints.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-set -euo pipefail
+set -euxo pipefail
 
 # -------------------------
 # Args / flags

.github/workflows/_test_backend.yml

@@ -0,0 +1,79 @@
+name: Test Backend
+
+on:
+  workflow_call:
+    inputs:
+      backend:
+        description: 'Backend to test (xnnpack, coreml, vulkan, qnn)'
+        required: true
+        type: string
+      flows:
+        description: 'JSON array of flows to test'
+        required: true
+        type: string
+      ref:
+        description: 'Git ref to checkout'
+        required: false
+        type: string
+        default: ${{ github.sha }}
+      timeout:
+        description: 'Job timeout in minutes'
+        required: false
+        type: number
+        default: 120
+      run-linux:
+        description: 'Whether to run Linux tests'
+        required: false
+        type: boolean
+        default: false
+      run-macos:
+        description: 'Whether to run macOS tests'
+        required: false
+        type: boolean
+        default: false
+
+jobs:
+  test-backend-linux:
+    if: ${{ inputs.run-linux }}
+    strategy:
+      fail-fast: false
+      matrix:
+        flow: ${{ fromJSON(inputs.flows) }}
+        suite: [models, operators]
+
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    with:
+      ref: ${{ inputs.ref }}
+      runner: linux.4xlarge.memory
+      docker-image: ci-image:executorch-ubuntu-22.04-clang12
+      submodules: recursive
+      timeout: ${{ inputs.timeout }}
+      upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
+      script: |
+        set -eux
+
+        source .ci/scripts/test_backend_linux.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
+
+  test-backend-macos:
+    if: ${{ inputs.run-macos }}
+    strategy:
+      fail-fast: false
+      matrix:
+        flow: ${{ fromJSON(inputs.flows) }}
+        suite: [models, operators]
+
+    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
+    with:
+      ref: ${{ inputs.ref }}
+      runner: macos-m1-stable
+      python-version: "3.12"
+      submodules: recursive
+      timeout: ${{ inputs.timeout }}
+      upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
+      script: |
+        set -eux
+
+        # This is needed to get the prebuilt PyTorch wheel from S3
+        ${CONDA_RUN} --no-capture-output pip install awscli==1.37.21
+
+        source .ci/scripts/test_backend_macos.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"

.github/workflows/nightly.yml

@@ -36,51 +36,3 @@ jobs:
     uses: ./.github/workflows/_link_check.yml
     with:
       ref: ${{ github.sha }}
-
-  backend-test-linux:
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    strategy:
-      fail-fast: false
-      matrix:
-        flow: [
-          qnn, qnn_16a16w, qnn_16a8w, qnn_16a4w, qnn_16a4w_block, qnn_8a8w,
-          vulkan, vulkan_static_int8_per_channel,
-          xnnpack, xnnpack_dynamic_int8_per_channel, xnnpack_static_int8_per_channel, xnnpack_static_int8_per_tensor
-        ]
-        suite: [models, operators]
-    with:
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      runner: linux.4xlarge.memory
-      docker-image: ci-image:executorch-ubuntu-22.04-clang12
-      submodules: recursive
-      timeout: 120
-      upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
-      script: |
-        set -eux
-
-        source .ci/scripts/test_backend_linux.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
-
-  backend-test-macos:
-    uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
-    permissions:
-      id-token: write
-      contents: read
-    strategy:
-      fail-fast: false
-      matrix:
-        flow: [coreml, coreml_static_int8]
-        suite: [models, operators]
-    with:
-      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      runner: macos-m1-stable
-      python-version: 3.12
-      submodules: recursive
-      timeout: 120
-      upload-artifact: test-report-${{ matrix.flow }}-${{ matrix.suite }}
-      script: |
-        set -eux
-
-        # This is needed to get the prebuilt PyTorch wheel from S3
-        ${CONDA_RUN} --no-capture-output pip install awscli==1.37.21
-
-        source .ci/scripts/test_backend_macos.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"

.github/workflows/pull.yml

@@ -779,6 +779,7 @@ jobs:
       contents: read
     strategy:
       fail-fast: false
+    if: false # TODO Re-enable after fixing timeouts (#14314)
     with:
       runner: linux.2xlarge
       docker-image: ci-image:executorch-ubuntu-22.04-gcc9

.github/workflows/test-backend-coreml.yml

@@ -0,0 +1,27 @@
+name: Test CoreML Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-coreml.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-coreml:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: coreml
+      flows: '["coreml", "coreml_static_int8"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-macos: true

.github/workflows/test-backend-qnn.yml

@@ -0,0 +1,27 @@
+name: Test QNN Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-qnn.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-qnn:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: qnn
+      flows: '["qnn", "qnn_16a16w", "qnn_16a8w", "qnn_16a4w", "qnn_16a4w_block", "qnn_8a8w"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true

.github/workflows/test-backend-vulkan.yml

@@ -0,0 +1,27 @@
+name: Test Vulkan Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-vulkan.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-vulkan:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: vulkan
+      flows: '["vulkan", "vulkan_static_int8_per_channel"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true

.github/workflows/test-backend-xnnpack.yml

@@ -0,0 +1,27 @@
+name: Test XNNPACK Backend
+
+on:
+  schedule:
+    - cron: 0 2 * * *
+  push:
+    tags:
+      - ciflow/nightly/*
+  pull_request:
+    paths:
+      - .github/workflows/test-backend-xnnpack.yml
+      - .github/workflows/_test_backend.yml
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}--${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+jobs:
+  test-xnnpack:
+    uses: ./.github/workflows/_test_backend.yml
+    with:
+      backend: xnnpack
+      flows: '["xnnpack", "xnnpack_dynamic_int8_per_channel", "xnnpack_static_int8_per_channel", "xnnpack_static_int8_per_tensor"]'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 120
+      run-linux: true

backends/arm/ethosu/backend.py

@@ -15,6 +15,7 @@
 from typing import final, List
 
 from executorch.backends.arm.arm_vela import vela_compile
+from executorch.backends.arm.ethosu.compile_spec import EthosUCompileSpec
 
 from executorch.backends.arm.tosa.backend import TOSABackend
 from executorch.exir.backend.backend_details import BackendDetails, PreprocessResult
@@ -35,16 +36,13 @@ class EthosUBackend(BackendDetails):
 
     @staticmethod
     def _compile_tosa_flatbuffer(
-        tosa_flatbuffer: bytes, compile_spec: List[CompileSpec]
+        tosa_flatbuffer: bytes, compile_spec: EthosUCompileSpec
     ) -> bytes:
         """
         Static helper method to do the compilation of the TOSA flatbuffer
         representation to a target specific binary stream.
         """
-        compile_flags = []
-        for spec in compile_spec:
-            if spec.key == "compile_flags":
-                compile_flags.append(spec.value.decode())
+        compile_flags = compile_spec.compiler_flags
 
         if len(compile_flags) == 0:
             # Not testing for compile_flags correctness here, just that they are
@@ -64,10 +62,11 @@ def _compile_tosa_flatbuffer(
     @staticmethod
     def preprocess(
         edge_program: ExportedProgram,
-        compile_spec: List[CompileSpec],
+        compile_specs: List[CompileSpec],
     ) -> PreprocessResult:
         logger.info(f"{EthosUBackend.__name__} preprocess")
 
+        compile_spec = EthosUCompileSpec.from_list(compile_specs)
         # deduce TOSA compile_spec from Ethos-U compile spec. We get a new
         # compile spec list, containing only elements relevant for the
         # TOSABackend.
@@ -77,7 +76,7 @@ def preprocess(
         # ('All backend implementation are final...'), so use composition instead.
         # preprocess returns the serialized TOSA flatbuffer in .processed_bytes,
         # which can be passed on to next compilation step.
-        tosa_preprocess = TOSABackend.preprocess(edge_program, tosa_compile_spec)
+        tosa_preprocess = TOSABackend._preprocess(edge_program, tosa_compile_spec)
 
         binary = EthosUBackend._compile_tosa_flatbuffer(
             tosa_preprocess.processed_bytes, compile_spec

backends/arm/operators/op_abs.py

@@ -73,7 +73,9 @@ def define_node(
             abs_output = output
 
         # Do the INT32 Abs
-        tosa_graph.addOperator(
+        self._serialize_operator(
+            node,
+            tosa_graph,
             ts.TosaOp.Op().ABS,
             [
                 rescaled_inputs[0].name,