Merge branch 'main' into ops

Author: zingo

build/build_android_instrumentation.sh renamed to .ci/scripts/build_android_instrumentation.sh

@@ -0,0 +1,11 @@
+# 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.
+
+# This file is sourced into the environment before building a pip wheel. It
+# should typically only contain shell variable assignments. Be sure to export
+# any variables so that subprocesses will see them.
+
+source "${GITHUB_WORKSPACE}/${REPOSITORY}/.ci/scripts/wheel/envvar_base.sh"

.ci/scripts/wheel/__init__.py

@@ -8,16 +8,7 @@
 # should typically only contain shell variable assignments. Be sure to export
 # any variables so that subprocesses will see them.
 
-# Enable pybindings so that users can execute ExecuTorch programs from python.
-export EXECUTORCH_BUILD_PYBIND=1
-
-# Ensure that CMAKE_ARGS is defined before referencing it. Defaults to empty
-# if not defined.
-export CMAKE_ARGS="${CMAKE_ARGS:-}"
-
-# Link the XNNPACK backend into the pybindings runtime so that users can execute
-# ExecuTorch programs that delegate to it.
-CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_XNNPACK=ON"
+source "${GITHUB_WORKSPACE}/${REPOSITORY}/.ci/scripts/wheel/envvar_base.sh"
 
 # When building for macOS, link additional backends into the pybindings runtime.
 CMAKE_ARGS="${CMAKE_ARGS} -DEXECUTORCH_BUILD_COREML=ON"

build/packaging/env_var_script_linux.sh renamed to .ci/scripts/wheel/envvar_base.sh

@@ -14,4 +14,4 @@ set -euxo pipefail
 # which does install them. Though we'd need to disable build isolation to be
 # able to see the installed torch package.
 
-pip install --progress-bar off -r requirements-dev.txt
+"${GITHUB_WORKSPACE}/${REPOSITORY}/install_requirements.sh"

.ci/scripts/wheel/envvar_linux.sh

@@ -0,0 +1,66 @@
+# 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.
+
+import os
+import subprocess
+import sys
+from dataclasses import dataclass
+from functools import cache
+from typing import List
+
+
+@cache
+def _unsafe_get_env(key: str) -> str:
+    value = os.getenv(key)
+    if value is None:
+        raise RuntimeError(f"environment variable '{key}' is not set")
+    return value
+
+
+@cache
+def _repository_root_dir() -> str:
+    return os.path.join(
+        _unsafe_get_env("GITHUB_WORKSPACE"),
+        _unsafe_get_env("REPOSITORY"),
+    )
+
+
+# For some reason, we are unable to see the entire repo in the python path.
+# So manually add it.
+sys.path.append(_repository_root_dir())
+from examples.models import Backend, Model
+
+
+@dataclass
+class ModelTest:
+    model: Model
+    backend: Backend
+
+
+def run_tests(model_tests: List[ModelTest]) -> None:
+    # Why are we doing this envvar shenanigans? Since we build the testers, which
+    # uses buck, we cannot run as root. This is a sneaky of getting around that
+    # test.
+    #
+    # This can be reverted if either:
+    #   - We remove usage of buck in our builds
+    #   - We stop running the Docker image as root: https://github.com/pytorch/test-infra/issues/5091
+    envvars = os.environ.copy()
+    envvars.pop("HOME")
+
+    for model_test in model_tests:
+        subprocess.run(
+            [
+                os.path.join(_repository_root_dir(), ".ci/scripts/test_model.sh"),
+                str(model_test.model),
+                # What to build `executor_runner` with for testing.
+                "cmake",
+                str(model_test.backend),
+            ],
+            env=envvars,
+            check=True,
+            cwd=_repository_root_dir(),
+        )

build/packaging/env_var_script_macos.sh renamed to .ci/scripts/wheel/envvar_macos.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env python
+# 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.
+
+import test_base
+from examples.models import Backend, Model
+
+if __name__ == "__main__":
+    test_base.run_tests(
+        model_tests=[
+            test_base.ModelTest(
+                model=Model.Mv3,
+                backend=Backend.XnnpackQuantizationDelegation,
+            )
+        ]
+    )

build/packaging/post_build_script.sh renamed to .ci/scripts/wheel/post_build_script.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env python
+# 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.
+
+import test_base
+from examples.models import Backend, Model
+
+if __name__ == "__main__":
+    test_base.run_tests(
+        model_tests=[
+            test_base.ModelTest(
+                model=Model.Mv3,
+                backend=Backend.XnnpackQuantizationDelegation,
+            )
+        ]
+    )

build/packaging/pre_build_script.sh renamed to .ci/scripts/wheel/pre_build_script.sh

@@ -31,7 +31,7 @@ jobs:
         # Build LLM Demo for Android
         export BUILD_AAR_DIR=aar-out
         bash build/build_android_library.sh ${ARTIFACTS_DIR_NAME}
-        bash build/build_android_instrumentation.sh ${ARTIFACTS_DIR_NAME}
+        bash .ci/scripts/build_android_instrumentation.sh ${ARTIFACTS_DIR_NAME}
 
         mkdir -p ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom
         bash ".ci/scripts/test_llama.sh" -model stories110M -build_tool cmake -dtype fp16 -mode portable -upload ${ARTIFACTS_DIR_NAME}/fp32-xnnpack-custom

.ci/scripts/wheel/test_base.py

@@ -4,7 +4,7 @@ name: Build Linux Wheels
 on:
   pull_request:
     paths:
-      - build/packaging/**
+      - .ci/**/*
       - .github/workflows/build-wheels-linux.yml
   push:
     branches:
@@ -39,9 +39,9 @@ jobs:
       matrix:
         include:
           - repository: pytorch/executorch
-            pre-script: build/packaging/pre_build_script.sh
-            post-script: build/packaging/post_build_script.sh
-            smoke-test-script: build/packaging/smoke_test.py
+            pre-script: .ci/scripts/wheel/pre_build_script.sh
+            post-script: .ci/scripts/wheel/post_build_script.sh
+            smoke-test-script: .ci/scripts/wheel/test_linux.py
             package-name: executorch
     name: ${{ matrix.repository }}
     uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main
@@ -51,11 +51,8 @@ jobs:
       test-infra-repository: pytorch/test-infra
       test-infra-ref: main
       build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
-      # ExecuTorch only needs the first layer of submodules; override the
-      # "recursive" default to do less work, and to give the buck daemon fewer
-      # files to look at.
-      submodules: true
-      env-var-script: build/packaging/env_var_script_linux.sh
+      submodules: recursive
+      env-var-script: .ci/scripts/wheel/envvar_linux.sh
       pre-script: ${{ matrix.pre-script }}
       post-script: ${{ matrix.post-script }}
       package-name: ${{ matrix.package-name }}

.ci/scripts/wheel/test_linux.py

@@ -4,7 +4,7 @@ name: Build macOS Wheels
 on:
   pull_request:
     paths:
-      - build/packaging/**
+      - .ci/**/*
       - .github/workflows/build-wheels-macos.yml
   push:
     branches:
@@ -39,9 +39,9 @@ jobs:
       matrix:
         include:
           - repository: pytorch/executorch
-            pre-script: build/packaging/pre_build_script.sh
-            post-script: build/packaging/post_build_script.sh
-            smoke-test-script: build/packaging/smoke_test.py
+            pre-script: .ci/scripts/wheel/pre_build_script.sh
+            post-script: .ci/scripts/wheel/post_build_script.sh
+            smoke-test-script: .ci/scripts/wheel/test_macos.py
             package-name: executorch
     name: ${{ matrix.repository }}
     uses: pytorch/test-infra/.github/workflows/build_wheels_macos.yml@main
@@ -51,12 +51,9 @@ jobs:
       test-infra-repository: pytorch/test-infra
       test-infra-ref: main
       build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
-      # ExecuTorch only needs the first layer of submodules; override the
-      # "recursive" default to do less work, and to give the buck daemon fewer
-      # files to look at.
-      submodules: true
+      submodules: recursive
       delocate-wheel: false
-      env-var-script: build/packaging/env_var_script_macos.sh
+      env-var-script: .ci/scripts/wheel/envvar_macos.sh
       pre-script: ${{ matrix.pre-script }}
       post-script: ${{ matrix.post-script }}
       package-name: ${{ matrix.package-name }}

.ci/scripts/wheel/test_macos.py

@@ -343,3 +343,29 @@ init_command = [
     '--dry-run={{DRYRUN}}',
     '--requirement=requirements-lintrunner.txt',
 ]
+
+[[linter]]
+code = 'LICENSELINT'
+include_patterns = [
+    '**/*',
+]
+exclude_patterns = [
+    '**/fb/**',
+    '.lintrunner.toml',
+]
+command = [
+    'python',
+    '-m',
+    'lintrunner_adapters',
+    'run',
+    'grep_linter',
+    '--pattern=Confidential and proprietary',
+    '--linter-name=LICENSELINT',
+    '--error-name=Wrong license',
+    """--error-description=\
+        Code contributed to ExecuTorch open source repo should have \
+        BSD-license header \
+    """,
+    '--',
+    '@{{PATHSFILE}}',
+]

.github/workflows/_android.yml

@@ -23,6 +23,7 @@
 
 # debug functionality
 logger = logging.getLogger(__name__)
+logger.setLevel(logging.WARNING)
 
 
 @final

.github/workflows/build-wheels-linux.yml

@@ -11,6 +11,7 @@
     pool_2d_support,
     reduce_sum_support,
     right_shift_support,
+    slice_copy_support,
     to_copy_support,
     tosa_supported_operators,
 )

.github/workflows/build-wheels-macos.yml

@@ -0,0 +1,39 @@
+# Copyright 2025 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 logging
+
+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.backends.arm.tosa_utils import getNodeArgs
+from executorch.exir.dialects._ops import ops as exir_ops
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.WARNING)
+
+
+@register_tosa_support_check
+class SliceCopySupported(SupportedTOSAOperatorCheck):
+    targets = [exir_ops.edge.aten.slice_copy.Tensor]
+
+    tosa_specs = [
+        TosaSpecification.create_from_string("TOSA-0.80+BI"),
+        TosaSpecification.create_from_string("TOSA-0.80+MI"),
+    ]
+
+    def is_node_tosa_supported(self, node: fx.Node, tosa_spec: TosaSpecification) -> bool:  # type: ignore[override, misc]
+        if tosa_spec not in self.tosa_specs:
+            return False
+
+        inputs = getNodeArgs(node)
+        if len(inputs) == 5 and (step := inputs[4].number) != 1:
+            logging.warning(f"{node.target} with step size of {step} not supported.")
+            return False
+        return True

.lintrunner.toml

@@ -75,7 +75,6 @@ def register_tosa_support_check(checker: Type[SupportedTOSAOperatorCheck]):
 def get_registered_tosa_support_checks(
     tosa_spec: TosaSpecification,
 ) -> list[Type[SupportedTOSAOperatorCheck]]:
-
     if tosa_spec not in _tosa_spec_support:
         raise RuntimeError(
             f"TOSA specification not valid: {tosa_spec} not in {list(_tosa_spec_support.keys())}"
@@ -165,7 +164,6 @@ def is_node_supported(
             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.tanh.default,
             exir_ops.edge.aten.upsample_nearest2d.vec,

backends/arm/ethosu_backend.py

@@ -45,6 +45,12 @@ def define_node(
         # Handle int8 (quantized) and int32
         assert inputs[0].dtype in [ts.DType.INT8, ts.DType.INT32]
 
+        dim_order = (
+            inputs[0].dim_order
+            if len(inputs[0].shape) > len(inputs[1].shape)
+            else inputs[1].dim_order
+        )
+
         if inputs[0].dtype == ts.DType.INT8:
             rescaled_inputs, scale_back = tqutils.insert_rescale_ops_to_int32(
                 tosa_graph, inputs, node
@@ -61,13 +67,14 @@ def define_node(
             # output.dtype == ts.DType.INT32
             add_output = output
 
+        input1, input2 = tutils.reshape_for_broadcast(
+            tosa_graph, rescaled_inputs, dim_order
+        )
+
         # Do the INT32 Add
         tosa_graph.addOperator(
             TosaOp.Op().ADD,
-            [
-                rescaled_inputs[0].name,
-                rescaled_inputs[1].name,
-            ],
+            [input1.name, input2.name],
             [add_output.name],
             None,
         )
@@ -108,10 +115,12 @@ def define_node(
             assert inputs[0].dtype == ts.DType.FP32
             assert output.dtype == ts.DType.FP32
 
+            input1, input2 = tutils.reshape_for_broadcast(tosa_graph, inputs)
+
             # MI lowering
             tosa_graph.addOperator(
                 TosaOp.Op().ADD,
-                [inputs[0].name, inputs[1].name],
+                [input1.name, input2.name],
                 [output.name],
                 None,
             )