BKC for Experimental Support on torch.compile for Windows on XPU

[Stonepia]

This document provides early experimental support for torch.compile on Windows with XPU. It tracks the status and known issues.

• 1. Overall Branch
• 2. Build Steps
  • 2.0.1. Windows Environment Setup
  • 2.0.2. Build PyTorch
  • 2.0.3. Build Triton
• 3. Running Setup

1. Overall Branch

Refer to the PyTorch PR: pytorch/pytorch#144303. Use the branch specified in that PR.
For Triton, use the branch: https://github.com/intel/intel-xpu-backend-for-triton/tree/hot-fixes-for-pytorch.

2. Build Steps

Currently, Triton needs to be built from source and installed. The PyTorch build process remains unchanged.

2.0.1. Windows Environment Setup

For more details about the env setting, please refer to this discussion.

1. Enable Long Path

# Enable long path for the system (Need admin)
New-ItemProperty -Path "HKLM:\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1 -PropertyType DWORD -Force

# Enable long path for git (Need admin)
git config --system core.longpaths true
git config --global core.longpaths true

2. Enable Symlink Creation
   Activate developer mode. It allows normal user to create symlinks, which may lead build failures during Triton build.

2.0.2. Build PyTorch

Use the branch in pytorch/pytorch#144303 . All the steps are the same with existing BKC.

2.0.3. Build Triton

Use the pinned commit in the above PR.

1. Download Level Zero SDK
   Please download level-zero-win-sdk-*.zip from https://github.com/oneapi-src/level-zero/releases. We tried with v1.19.2.
   Unzip the file and put the folder to some path like C:\level_zero.

2. Build Triton
   Open the Intel oneAPI command prompt for Intel 64 for Visual Studio 2022 or activate oneAPI env by:

"C:\Program Files (x86)\Intel\oneAPI\<toolkit-version>\oneapi-vars.bat"

Set the following env flag for Triton build:

set VS2022INSTALLDIR="C:\Program Files\Microsoft Visual Studio\2022\Community"
set ZE_PATH=C:\level_zero
set CL=/D_CRT_SECURE_NO_WARNINGS
"C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvars64.bat"
"C:\Program Files (x86)\Intel\oneAPI\2025.0\oneapi-vars.bat"

Build Triton. Please put the triton folder in a shallow folder path (e.g., C:\triton)

git clone https://github.com/intel/intel-xpu-backend-for-triton triton
cd triton
git checkout c23ff25775780cc4bb1ca530fd3ae33b0cf3b56e

cd python
pip install -U wheel pybind11 certifi cython cmake setuptools>=65.6.1
python -m certifi
pip install -v --no-build-isolation '.[build,tests,tutorials]'

One can also use python setup.py bdist_wheels in triton\python to get the wheels. Then pip install dist\*.whl.

3. Running Setup

The overall running setup is the same. One additional step is to be sure to add level-zero to ZE_PATH:

set ZE_PATH=C:\level_zero

Then one could run the tests. Before running tests (especially PyTorch UT), please clean the TEMP folder to reduce size.
Due to the limitations of Windows OS, the TEMP may not be cleaned automatically.

del /q %TEMP%\*  & rd /s /q %TEMP%

Then you could run like below:

cd pytorch\test\inductor
pytest -v -k xpu test_torchinductor.py