@tinglvv is attempting to deploy a commit to the Meta Open Source Team on Vercel.

A member of the Team first needs to authorize it.

Deprecating CUDA_CUDDN_VERSIONS dict as I don't see it called elsewhere (the name also had a typo in it)

Deprecating CUDA_CUDDN_VERSIONS dict as I don't see it called elsewhere (the name also had a typo in it)

Referenced by: https://github.com/pytorch/test-infra/blob/main/tools/scripts/generate_docker_release_matrix.py#L48

Deprecating CUDA_CUDDN_VERSIONS dict as I don't see it called elsewhere (the name also had a typo in it)

Referenced by: https://github.com/pytorch/test-infra/blob/main/tools/scripts/generate_docker_release_matrix.py#L48

Thanks! Did a global search and found it. Let me add it back.

@atalman

What's the policy on how many versions of CUDA we need to test for?

(11.8, 12.4, 12.6, and now 12.8 would be 4 versions, is there a process where we go back and remove some of the tests?)

Same with Python do we really need to test on 5 versions of Python 3.9, 3.10, 3.11, 3.12, and 3.13?

@atalman

What's the policy on how many versions of CUDA we need to test for?

(11.8, 12.4, 12.6, and now 12.8 would be 4 versions, is there a process where we go back and remove some of the tests?)

Same with Python do we really need to test on 5 versions of Python 3.9, 3.10, 3.11, 3.12, and 3.13?

We discussed this on the CI Sync meeting today and @atalman shared with us a link (https://github.com/pytorch/rfcs/blob/master/RFC-0038-cpython-support.md). Looks like we support every released version of Python as long as its not EOL.

For CUDA 3 versions are supported so 12.4 will be removed once 12.8 is confirmed working. This answers my questions. Thanks!

Hi @zxiiro For each cuda update we have RFC issue: pytorch/pytorch#145544 where release matrix is discussed.

Just noticed we should skip only the Windows build for 12.8, fix in #6252