[Windows] Add CI wheel build jobs for Windows

.ci/scripts/test_model.ps1

@@ -34,7 +34,7 @@ function ExportModel-Xnnpack {
         [bool]$quantize
     )
 
-    if $(quantize) {
+    if ($quantize) {
         python -m examples.xnnpack.aot_compiler --model_name="${MODEL_NAME}" --delegate --quantize | Write-Host
         $modelFile = "$($modelName)_xnnpack_q8.pte"
     } else {

.ci/scripts/wheel/pre_build_script.sh

@@ -9,9 +9,26 @@ set -euxo pipefail
 
 # This script is run before building ExecuTorch binaries
 
+# Clone nested submodules for tokenizers - this is a workaround for recursive
+# submodule clone failing due to path length limitations on Windows. Eventually,
+# we should update the core job in test-infra to enable long paths before
+# checkout to avoid needing to do this.
+pushd extension/llm/tokenizers
+git submodule update --init
+popd
+
+# On Windows, enable symlinks and re-checkout the current revision to create
+# the symlinked src/ directory. This is needed to build the wheel.
+UNAME_S=$(uname -s)
+if [[ $UNAME_S == *"MINGW"* || $UNAME_S == *"MSYS"* ]]; then
+    echo "Enabling symlinks on Windows"
+    git config core.symlinks true
+    git checkout -f HEAD
+fi
+
 # Manually install build requirements because `python setup.py bdist_wheel` does
 # not install them. TODO(dbort): Switch to using `python -m build --wheel`,
 # which does install them. Though we'd need to disable build isolation to be
 # able to see the installed torch package.
 
-"${GITHUB_WORKSPACE}/${REPOSITORY}/install_requirements.sh"  --example
+"${GITHUB_WORKSPACE}/${REPOSITORY}/install_requirements.sh" --example

.ci/scripts/wheel/test_windows.py

@@ -0,0 +1,75 @@
+#!/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.
+
+from typing import List
+
+import torch
+from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner
+from executorch.examples.models import Backend, Model, MODEL_NAME_TO_MODEL
+from executorch.examples.models.model_factory import EagerModelFactory
+from executorch.examples.xnnpack import MODEL_NAME_TO_OPTIONS
+from executorch.examples.xnnpack.quantization.utils import quantize as quantize_xnn
+from executorch.exir import EdgeCompileConfig, to_edge_transform_and_lower
+from executorch.extension.pybindings.portable_lib import (
+    _load_for_executorch_from_buffer,
+)
+from test_base import ModelTest
+
+
+def test_model_xnnpack(model: Model, quantize: bool) -> None:
+    model_instance, example_inputs, _, _ = EagerModelFactory.create_model(
+        *MODEL_NAME_TO_MODEL[str(model)]
+    )
+
+    model_instance.eval()
+    ref_outputs = model_instance(*example_inputs)
+
+    if quantize:
+        quant_type = MODEL_NAME_TO_OPTIONS[str(model)].quantization
+        model_instance = torch.export.export_for_training(
+            model_instance, example_inputs
+        )
+        model_instance = quantize_xnn(
+            model_instance.module(), example_inputs, quant_type
+        )
+
+    lowered = to_edge_transform_and_lower(
+        torch.export.export(model_instance, example_inputs),
+        partitioner=[XnnpackPartitioner()],
+        compile_config=EdgeCompileConfig(
+            _check_ir_validity=False,
+        ),
+    ).to_executorch()
+
+    loaded_model = _load_for_executorch_from_buffer(lowered.buffer)
+    et_outputs = loaded_model([*example_inputs])
+
+    if isinstance(ref_outputs, torch.Tensor):
+        ref_outputs = (ref_outputs,)
+
+    assert len(ref_outputs) == len(et_outputs)
+    for i in range(len(ref_outputs)):
+        torch.testing.assert_close(ref_outputs[i], et_outputs[i], atol=1e-4, rtol=1e-5)
+
+
+def run_tests(model_tests: List[ModelTest]) -> None:
+    for model_test in model_tests:
+        if model_test.backend == Backend.Xnnpack:
+            test_model_xnnpack(model_test.model, quantize=False)
+        else:
+            raise RuntimeError(f"Unsupported backend {model_test.backend}.")
+
+
+if __name__ == "__main__":
+    run_tests(
+        model_tests=[
+            ModelTest(
+                model=Model.Mv3,
+                backend=Backend.Xnnpack,
+            ),
+        ]
+    )

.ci/scripts/wheel/vc_env_helper.bat

@@ -0,0 +1,61 @@
+REM This is lightly modified from the torchvision Windows build logic.
+REM See https://github.com/pytorch/vision/blob/main/packaging/windows/internal/vc_env_helper.bat
+
+@echo on
+
+set VC_VERSION_LOWER=17
+set VC_VERSION_UPPER=18
+
+for /f "usebackq tokens=*" %%i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe" -legacy -products * -version [%VC_VERSION_LOWER%^,%VC_VERSION_UPPER%^) -property installationPath`) do (
+    if exist "%%i" if exist "%%i\VC\Auxiliary\Build\vcvarsall.bat" (
+        set "VS15INSTALLDIR=%%i"
+        set "VS15VCVARSALL=%%i\VC\Auxiliary\Build\vcvarsall.bat"
+        goto vswhere
+    )
+)
+
+:vswhere
+if "%VSDEVCMD_ARGS%" == "" (
+    call "%VS15VCVARSALL%" x64 || exit /b 1
+) else (
+    call "%VS15VCVARSALL%" x64 %VSDEVCMD_ARGS% || exit /b 1
+)
+
+@echo on
+
+if "%CU_VERSION%" == "xpu" call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
+
+set DISTUTILS_USE_SDK=1
+
+set args=%1
+shift
+:start
+if [%1] == [] goto done
+set args=%args% %1
+shift
+goto start
+
+:done
+if "%args%" == "" (
+    echo Usage: vc_env_helper.bat [command] [args]
+    echo e.g. vc_env_helper.bat cl /c test.cpp
+)
+
+set work_dir=%CD%
+if exist setup.py (
+    echo "Creating symlink..."
+    REM Setup a symlink to shorten the path length.
+    REM Note that the ET directory has to be named "executorch".
+    cd %GITHUB_WORKSPACE%
+    if not exist et\ (
+        mkdir et
+    )
+    cd et
+    echo Work dir: %work_dir%
+    if not exist executorch\ (
+        mklink /d executorch %work_dir%
+    )
+    cd executorch
+)
+
+%args% || exit /b 1

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

@@ -0,0 +1,59 @@
+name: Build Windows Wheels
+
+on:
+  pull_request:
+  push:
+    branches:
+      - nightly
+      - main
+      - release/*
+    tags:
+        # NOTE: Binary build pipelines should only get triggered on release candidate builds
+        # Release candidate tags look like: v1.11.0-rc1
+        - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
+  workflow_dispatch:
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+  generate-matrix:
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    with:
+      package-type: wheel
+      os: windows
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      with-cuda: disabled
+      with-rocm: disabled
+      python-versions: '["3.10", "3.11", "3.12"]'
+
+  build:
+    needs: generate-matrix
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - repository: pytorch/executorch
+            pre-script: .ci\\scripts\\wheel\\pre_build_script.sh
+            env-script: .ci\\scripts\\wheel\\vc_env_helper.bat
+            post-script: .ci\\scripts\\wheel\\post_build_script.sh
+            smoke-test-script: .ci/scripts/wheel/test_windows.py
+            package-name: executorch
+    name: ${{ matrix.repository }}
+    uses: pytorch/test-infra/.github/workflows/build_wheels_windows.yml@main
+    with:
+      repository: ${{ matrix.repository }}
+      ref: ""
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
+      pre-script: ${{ matrix.pre-script }}
+      env-script: ${{ matrix.env-script }}
+      post-script: ${{ matrix.post-script }}
+      package-name: ${{ matrix.package-name }}
+      smoke-test-script: ${{ matrix.smoke-test-script }}
+      trigger-event: ${{ github.event_name }}
+      wheel-build-params: "--verbose"
+      submodules: true

examples/models/__init__.py

@@ -44,6 +44,7 @@ def __str__(self) -> str:
 
 
 class Backend(str, Enum):
+    Xnnpack = "xnnpack"
     XnnpackQuantizationDelegation = "xnnpack-quantization-delegation"
     CoreMlExportOnly = "coreml"
     CoreMlExportAndTest = "coreml-test"  # AOT export + test with runner

setup.py

@@ -142,6 +142,7 @@ def string(cls) -> str:
     @classmethod
     def write_to_python_file(cls, path: str) -> None:
         """Creates a file similar to PyTorch core's `torch/version.py`."""
+
         lines = [
             "from typing import Optional",
             '__all__ = ["__version__", "git_version"]',
@@ -773,7 +774,7 @@ def run(self):  # noqa C901
     # platform-specific files using InstallerBuildExt.
     ext_modules=[
         BuiltFile(
-            src_dir="%CMAKE_CACHE_DIR%/third-party/flatc_proj/bin/",
+            src_dir="%CMAKE_CACHE_DIR%/third-party/flatc_ep/bin/",
             src_name="flatc",
             dst="executorch/data/bin/",
             is_executable=True,

third-party/CMakeLists.txt

@@ -29,7 +29,8 @@ endif()
 # Otherwise, flatc will target the project's toolchain (i.e. iOS, or Android).
 ExternalProject_Add(
   flatbuffers_ep
-  PREFIX ${CMAKE_CURRENT_BINARY_DIR}/flatc_proj
+  PREFIX ${CMAKE_CURRENT_BINARY_DIR}/flatc_ep
+  BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR}/flatc_ep/src/build
   SOURCE_DIR ${PROJECT_SOURCE_DIR}/third-party/flatbuffers
   CMAKE_ARGS -DFLATBUFFERS_BUILD_FLATC=ON
              -DFLATBUFFERS_INSTALL=ON
@@ -82,6 +83,7 @@ ExternalProject_Add(
   flatcc_ep
   PREFIX ${CMAKE_CURRENT_BINARY_DIR}/flatcc_ep
   SOURCE_DIR ${PROJECT_SOURCE_DIR}/third-party/flatcc
+  BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR}/flatcc_ep/src/build
   CMAKE_ARGS -DFLATCC_RTONLY=OFF
              -DFLATCC_TEST=OFF
              -DFLATCC_REFLECTION=OFF