Use the same CUDNN version on Windows as PyTorch

[huydhn]

11.7 uses 8.5.0; 11.8 uses 8.7.0; 12.1 uses 8.8.1. Otherwise, Windows vision job (8.5.0) would overwrite the CUDNN version setup by PyTorch (8.7.0) leading to this flaky failures https://github.com/pytorch/pytorch/actions/runs/5088860652/jobs/9146641450

RuntimeError: cuDNN version incompatibility: PyTorch was compiled  against (8, 7, 0) but found runtime version (8, 5, 0). PyTorch already comes bundled with cuDNN. One option to resolving this error is to ensure PyTorch can find the bundled cuDNN.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/7635

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure, 1 Pending

As of commit 2441e59:

NEW FAILURE - The following job has failed:

• pytorch/vision / conda-py3_8-cuda12_1 (gh)

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

[pmeier]

The only time we are hitting this script is here:

vision/.github/workflows/build-cmake.yml

Lines 79 to 86 in 81d12c3

	# FIXME: Basically, we are reinstalling CUDA here. We only need this, because we need to copy some files that
	# can be extracted from the CUDA installer, but are not available on our Windows AMI.
	# See https://github.com/pytorch/test-infra/pull/4189
	if [[ ${{ matrix.gpu-arch-type }} == cuda ]]; then
	export CU_VERSION=cu$(echo ${{ matrix.gpu-arch-version }} | sed 's/\.//')
	echo CU_VERSION="${CU_VERSION}"
	packaging/windows/internal/cuda_install.bat
	fi

@atalman and I are in the process to act on the FIXME so that TorchVision can just rely on the files already present in the AMI. See pytorch/test-infra#4227.

Stamp to unblock in case that takes too long.

[huydhn]

The only time we are hitting this script is here:

vision/.github/workflows/build-cmake.yml

Lines 79 to 86 in 81d12c3

	# FIXME: Basically, we are reinstalling CUDA here. We only need this, because we need to copy some files that
	# can be extracted from the CUDA installer, but are not available on our Windows AMI.
	# See https://github.com/pytorch/test-infra/pull/4189
	if [[ ${{ matrix.gpu-arch-type }} == cuda ]]; then
	export CU_VERSION=cu$(echo ${{ matrix.gpu-arch-version }} | sed 's/\.//')
	echo CU_VERSION="${CU_VERSION}"
	packaging/windows/internal/cuda_install.bat
	fi

@atalman and I are in the process to act on the FIXME so that TorchVision can just rely on the files already present in the AMI. See pytorch/test-infra#4227.

Stamp to unblock in case that takes too long.

Thank you for the context. It's indeed nicer to not need to do the installation step during CI.

[github-actions]

Hey @huydhn!

You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py