2024-11-18 nightly release (8526d0a)

Author: pytorchbot

.ci/scripts/gather_test_models.py

@@ -25,6 +25,7 @@
         "resnet50": "linux.12xlarge",
         "llava": "linux.12xlarge",
         "llama3_2_vision_encoder": "linux.12xlarge",
+        # "llama3_2_text_decoder": "linux.12xlarge",  # TODO: re-enable test when Huy's change is in / model gets smaller.
         # This one causes timeout on smaller runner, the root cause is unclear (T161064121)
         "dl3": "linux.12xlarge",
         "emformer_join": "linux.12xlarge",

.ci/scripts/test_model.sh

@@ -87,8 +87,8 @@ test_model() {
     bash examples/models/llava/install_requirements.sh
     STRICT="--no-strict"
   fi
-  if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" ]]; then
-    # Install requirements for llama vision
+  if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" || "$MODEL_NAME" == "llama3_2_text_decoder" ]]; then
+    # Install requirements for llama vision.
     bash examples/models/llama3_2_vision/install_requirements.sh
   fi
   # python3 -m examples.portable.scripts.export --model_name="llama2" should works too

.github/workflows/apple.yml

@@ -20,16 +20,28 @@ on:
       - extension/benchmark/apple/**
       - extension/module/**
   workflow_dispatch:
+  schedule:
+    - cron: '0 10 * * *'  # Runs daily at 2 AM PST
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
   cancel-in-progress: true
 
 jobs:
+  set-version:
+    runs-on: ubuntu-22.04
+    outputs:
+      version: ${{ steps.set_version.outputs.version }}
+    steps:
+      - name: Set VERSION variable
+        id: set_version
+        shell: bash
+        run: |
+          VERSION="0.4.0.$(TZ='PST8PDT' date +%Y%m%d)"
+          echo "version=$VERSION" >> "$GITHUB_OUTPUT"
+
   build-demo-ios:
     name: build-demo-ios
-    # NB: Don't run this on fork PRs because they won't have access to the secret and would fail anyway
-    if: ${{ !github.event.pull_request.head.repo.fork }}
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     secrets: inherit
     with:
@@ -41,6 +53,8 @@ jobs:
       secrets-env: BUILD_CERTIFICATE_BASE64 EXECUTORCH_DEMO_BUILD_PROVISION_PROFILE_BASE64 KEYCHAIN_PASSWORD
       upload-artifact: ios-apps
       script: |
+        set -eux
+
         BUILD_TOOL=cmake
 
         .ci/scripts/setup-conda.sh
@@ -59,7 +73,7 @@ jobs:
 
         # Build and test iOS Demo App
         PYTHON_EXECUTABLE=python ${CONDA_RUN} --no-capture-output \
-        build/test_ios_ci.sh ${ARTIFACTS_DIR_NAME}
+        build/test_ios_ci.sh "${ARTIFACTS_DIR_NAME}"
 
   # Upload the test demo app to S3
   upload-demo-ios:
@@ -77,6 +91,7 @@ jobs:
         shell: bash
         working-directory: ${{ runner.temp }}/artifacts/
         run: |
+          set -eux
           ls -lah ./
 
       - name: Upload the artifacts to S3
@@ -114,6 +129,7 @@ jobs:
 
   build-frameworks-ios:
     name: build-frameworks-ios
+    needs: set-version
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     with:
       runner: macos-latest-xlarge
@@ -123,8 +139,10 @@ jobs:
       upload-artifact: executorch-frameworks-ios
       timeout: 90
       script: |
+        set -eux
+
         BUILD_TOOL=cmake
-        VERSION="latest"
+        VERSION="${{ needs.set-version.outputs.version }}"
         FRAMEWORKS=(
           "executorch"
           "backend_coreml"
@@ -173,13 +191,17 @@ jobs:
 
   upload-frameworks-ios:
     runs-on: ubuntu-22.04
-    needs: build-frameworks-ios
+    needs: [build-frameworks-ios, set-version]
     timeout-minutes: 30
+    environment: ${{ github.ref == 'refs/heads/main' && 'cherry-pick-bot' || '' }}
     permissions:
       id-token: write
-      contents: read
+      contents: write
     steps:
       - uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+          token: ${{ secrets.GH_PYTORCHBOT_CHERRY_PICK_TOKEN || secrets.GITHUB_TOKEN }}
       - uses: actions/setup-python@v4
         with:
           python-version: '3.11'
@@ -196,15 +218,15 @@ jobs:
           name: executorch-frameworks-ios
           path: ${{ runner.temp }}/frameworks-ios/
       - name: Only push to S3 when running the workflow manually from main branch
-        if: ${{ github.event_name == 'workflow_dispatch' && github.ref == 'refs/heads/main' }}
+        if: ${{ (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && github.ref == 'refs/heads/main' }}
         shell: bash
         run: |
-          set -eux
           echo "UPLOAD_ON_MAIN=1" >> "${GITHUB_ENV}"
       - name: Upload the artifact to ossci-ios S3 bucket
         shell: bash
         run: |
           set -eux
+          VERSION="${{ needs.set-version.outputs.version }}"
 
           pip install awscli==1.32.18
 
@@ -215,14 +237,47 @@ jobs:
 
           for FILENAME in "${RUNNER_TEMP}"/frameworks-ios/*.zip; do
             [ -e "${FILENAME}" ] || continue
-            shasum -a 256 "${FILENAME}"
+            FRAMEWORK_NAME=$(basename "${FILENAME}" | sed "s/-${VERSION}.zip//")
+            CHECKSUM=$(shasum -a 256 "${FILENAME}" | cut -d ' ' -f1)
+            echo "${FRAMEWORK_NAME} ${CHECKSUM}" >> "${RUNNER_TEMP}/checksums.txt"
             ${AWS_CMD} "${FILENAME}" s3://ossci-ios/executorch/ --acl public-read
           done
+      - name: Update SwiftPM
+        shell: bash
+        run: |
+          set -eux
+          VERSION="${{ needs.set-version.outputs.version }}"
+          BRANCH="swiftpm-${VERSION}"
+
+          git checkout swiftpm
+
+          if git show-ref --verify --quiet refs/heads/${BRANCH}; then
+            git checkout "${BRANCH}"
+          else
+            git checkout -b "${BRANCH}"
+          fi
+
+          [[ -f Package.swift ]] || mv Package.swift.template Package.swift
+
+          sed -i "s/__VERSION__/${VERSION}/g" Package.swift
+
+          while read -r FRAMEWORK CHECKSUM; do
+            sed -i "s/__SHA256_${FRAMEWORK}__/${CHECKSUM}/g" Package.swift
+          done < "${RUNNER_TEMP}/checksums.txt"
+
+          if [[ "${UPLOAD_ON_MAIN:-0}" == "1" ]]; then
+            git config --global user.name "PyTorch Bot"
+            git config --global user.email "[email protected]"
+            git add Package.swift
+            git commit -am "${VERSION}"
+            git push -f origin "${BRANCH}"
+          else
+            echo "Draft Package.swift:"
+            cat Package.swift
+          fi
 
   build-benchmark-app:
     name: build-benchmark-app
-    # NB: Don't run this on fork PRs because they won't have access to the secret and would fail anyway
-    if: ${{ !github.event.pull_request.head.repo.fork }}
     uses: pytorch/test-infra/.github/workflows/macos_job.yml@main
     secrets: inherit
     with:
@@ -285,5 +340,5 @@ jobs:
         echo "::group::Build ExecuTorch benchmark app"
         mkdir -p extension/benchmark/apple/Benchmark/Models
         ${CONDA_RUN} --no-capture-output \
-          build/build_apple_llm_demo.sh ${ARTIFACTS_DIR_NAME}
+          build/build_apple_llm_demo.sh "${ARTIFACTS_DIR_NAME}"
         echo "::endgroup::"

.github/workflows/trunk.yml

@@ -411,8 +411,6 @@ jobs:
         pip install -U "huggingface_hub[cli]"
         huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
         pip install accelerate sentencepiece
-        # TODO(guangyang): Switch to use released transformers library after all required patches are included
-        pip install "git+https://github.com/huggingface/transformers.git@6cc4dfe3f1e8d421c6d6351388e06e9b123cbfe1"
         pip list
         echo "::endgroup::"
 

backends/arm/TARGETS

@@ -9,6 +9,7 @@ python_library(
     typing = True,
     deps = [
         ":arm_backend",
+        "//executorch/backends/arm/operator_support:operator_support",
         "//executorch/backends/arm/_passes:passes",
         "//executorch/exir:lib",
     ],

backends/arm/arm_partitioner.py

@@ -6,27 +6,27 @@
 # pyre-unsafe
 
 import logging
-import operator
 import os
-from typing import Callable, cast, final, List, Optional, Tuple
+from typing import Callable, final, List, Optional, Tuple
 
 import torch
 from executorch.backends.arm.arm_backend import ArmBackend  # usort: skip
 from executorch.backends.arm._passes.tag_io_quant_pass import TagIOQuantPass
+from executorch.backends.arm.operator_support.tosa_supported_operators import (
+    TOSASupportedOperators,
+)
+from executorch.backends.arm.tosa_specification import TosaSpecification
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 from executorch.exir.backend.partitioner import (
     DelegationSpec,
     Partitioner,
     PartitionResult,
 )
 from executorch.exir.backend.utils import tag_constant_data
-from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.passes import PassManager
 from torch.export.exported_program import ExportedProgram
 from torch.fx.passes.infra.partitioner import CapabilityBasedPartitioner
 
-from torch.fx.passes.operator_support import OperatorSupportBase
-
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.WARNING)
 TOSA_DBG_VERBOSE = os.environ.get("TOSA_DBG_VERBOSE") == "1"
@@ -35,71 +35,6 @@
     logger.setLevel(logging.INFO)
 
 
-class TOSASupportedOperators(OperatorSupportBase):
-    def is_node_supported(self, submodules, node: torch.fx.Node) -> bool:
-        supported = node.op == "call_function" and node.target in [
-            exir_ops.edge.aten.add.Tensor,
-            exir_ops.edge.aten.expand_copy.default,
-            exir_ops.edge.aten.cat.default,
-            exir_ops.edge.aten.bmm.default,
-            exir_ops.edge.aten.permute_copy.default,
-            exir_ops.edge.aten.hardtanh.default,
-            exir_ops.edge.aten.convolution.default,
-            exir_ops.edge.aten.div.Tensor,
-            exir_ops.edge.aten.exp.default,
-            exir_ops.edge.aten.log.default,
-            exir_ops.edge.aten.linear.default,
-            exir_ops.edge.aten.split_with_sizes_copy.default,
-            exir_ops.edge.aten.full.default,
-            exir_ops.edge.aten.mul.Tensor,
-            exir_ops.edge.aten._native_batch_norm_legit_no_training.default,
-            exir_ops.edge.aten.native_layer_norm.default,
-            exir_ops.edge.aten.avg_pool2d.default,
-            exir_ops.edge.aten.max_pool2d_with_indices.default,
-            exir_ops.edge.aten.sigmoid.default,
-            exir_ops.edge.aten.mm.default,
-            exir_ops.edge.aten.repeat.default,
-            exir_ops.edge.aten.reciprocal.default,
-            exir_ops.edge.aten.relu.default,
-            exir_ops.edge.aten.rsqrt.default,
-            exir_ops.edge.aten._softmax.default,
-            exir_ops.edge.aten.select_copy.int,
-            exir_ops.edge.aten._log_softmax.default,
-            exir_ops.edge.aten.slice_copy.Tensor,
-            exir_ops.edge.aten.sub.Tensor,
-            exir_ops.edge.aten.sum.dim_IntList,
-            exir_ops.edge.aten.tanh.default,
-            exir_ops.edge.aten.upsample_nearest2d.vec,
-            exir_ops.edge.aten.view_copy.default,
-            exir_ops.edge.aten.clone.default,
-            exir_ops.edge.aten.mean.dim,
-            exir_ops.edge.aten.var.correction,
-            exir_ops.edge.aten.unsqueeze_copy.default,
-            exir_ops.edge.aten.squeeze_copy.dims,
-            operator.getitem,
-            exir_ops.edge.quantized_decomposed.quantize_per_tensor.default,
-            exir_ops.edge.quantized_decomposed.dequantize_per_tensor.default,
-        ]
-
-        supported &= self.is_node_supported_custom(node)
-
-        # Override partitioning based on pre partition passes
-        if "arm_override_partition" in node.meta:
-            supported = supported & node.meta["arm_override_partition"]
-            node.meta.pop("arm_override_partition")
-
-        return supported
-
-    def is_node_supported_custom(self, node: torch.fx.Node) -> bool:
-        if node.target == exir_ops.edge.aten.mean.dim:
-            keep_dim = node.args[2] if len(node.args) > 2 else False
-            return cast(bool, keep_dim)
-        if node.target == exir_ops.edge.aten.var.correction:
-            keep_dim = node.kwargs.get("keepdim", False)
-            return cast(bool, keep_dim)
-        return True
-
-
 @final
 class ArmPartitioner(Partitioner):
     def __init__(self, compile_spec: List[CompileSpec]) -> None:
@@ -111,6 +46,12 @@ def partition(self, exported_program: ExportedProgram) -> PartitionResult:
         logger.info("ArmPartitioner::partition")
         partition_tags = {}
 
+        tosa_spec = TosaSpecification.create_from_compilespecs(
+            self.delegation_spec.compile_specs
+        )
+
+        logger.info(f"Partitioning for {tosa_spec}")
+
         for spec in self.delegation_spec.compile_specs:
             if spec.key == "quantize_io" and spec.value.decode() == "True":
                 # Exclude IO quantization from the partition
@@ -123,7 +64,7 @@ def partition(self, exported_program: ExportedProgram) -> PartitionResult:
 
         capability_partitioner = CapabilityBasedPartitioner(
             exported_program.graph_module,
-            TOSASupportedOperators(),
+            TOSASupportedOperators(tosa_spec),
             allows_single_node_partition=True,
         )
         partition_list = capability_partitioner.propose_partitions()

backends/arm/operator_support/TARGETS

@@ -0,0 +1,12 @@
+load("@fbcode_macros//build_defs:python_library.bzl", "python_library")
+
+python_library(
+    name = "operator_support",
+    srcs = glob(["*.py"]),
+    typing = True,
+    deps = [
+        "//executorch/backends/xnnpack/_passes:xnnpack_passes",
+        "//executorch/exir:lib",
+        "//executorch/backends/arm:tosa_specification"
+    ],
+)

backends/arm/operator_support/__init__.py

@@ -0,0 +1,8 @@
+# Copyright 2024 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.
+
+# pyre-unsafe
+
+from . import mean_dim_support, tosa_supported_operators, var_correction_support  # noqa

backends/arm/operator_support/mean_dim_support.py

@@ -0,0 +1,33 @@
+# Copyright 2024 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.
+
+# pyre-unsafe
+
+from typing import cast
+
+import torch.fx as fx
+
+from executorch.backends.arm.operator_support.tosa_supported_operators import (
+    register_tosa_support_check,
+    SupportedTOSAOperatorCheck,
+)
+from executorch.backends.arm.tosa_specification import TosaSpecification
+from executorch.exir.dialects._ops import ops as exir_ops
+
+
+@register_tosa_support_check
+class MeanDimSupported(SupportedTOSAOperatorCheck):
+    targets = [exir_ops.edge.aten.mean.dim]
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80.0+BI"),
+        TosaSpecification.create_from_string("TOSA-0.80.0+MI"),
+    ]
+
+    def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification) -> bool:
+        assert node.target in self.targets
+
+        keep_dim = node.args[2] if len(node.args) > 2 else False
+        return cast(bool, keep_dim)