Create build_wheels_windows.yml

[msaroufim]

Mostly copied from the torchvision workflows, upload working fine, need to test for real and will validate on my PC

Seems like it's working fine although I'm having issues building custom cuda extensions on Windows for cuda 11.8 only

• FAILED: C:/actions-runner/_work/ao/ao/pytorch/ao/build/temp.win-amd64-cpython-39/Release/torchao/csrc/cuda/sparse_marlin/marlin_kernel_nm.obj
• FAILED: C:/actions-runner/_work/ao/ao/pytorch/ao/build/temp.win-amd64-cpython-39/Release/torchao/csrc/cuda/tensor_core_tiled_layout/tensor_core_tiled_layout.obj
• FAILED: C:/actions-runner/_work/ao/ao/pytorch/ao/build/temp.win-amd64-cpython-39/Release/torchao/csrc/cuda/fp6_llm/fp6_linear.obj

So I tried to instead skip cuda 11.8 builds by adding the below, unfortunately the nova builds don't allow me to configure a specific cuda version but i suspect the below will just make build success but upload fail

REM Check CUDA version and skip build if it's 11.8
if "%CUDA_VERSION%" == "11.8" (
    echo Skipping build for CUDA 11.8
    exit /b 0
)

REM Set TORCH_CUDA_ARCH_LIST
set TORCH_CUDA_ARCH_LIST=8.0;8.6
if "%CU_VERSION%" == "cu124" set TORCH_CUDA_ARCH_LIST=%TORCH_CUDA_ARCH_LIST%;9.0

EDIT: Well considering cuda extensions arent build anywhere on windows and triton windows support is still experimental it seems fine as a stop gap to only publish windows cpu binaries for now

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1101

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 9d6c5d3 with merge base 6b52996 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[gau-nernst]

The undefined symbols are probably caused by building for earlier compute capabilities. Something like this would fix it.

ao/packaging/env_var_script_linux.sh

Lines 14 to 19 in 6b52996

	# Set ARCH list so that we can build fp16 with SM75+, the logic is copied from
	# pytorch/builder
	TORCH_CUDA_ARCH_LIST="8.0;8.6"
	if [[ ${CU_VERSION:-} == "cu124" ]]; then
	TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
	fi

The CUDA sources are actually not being compiled in 12.1 and 12.4 job, so that is kinda strange.