Skip to content

Conversation

NicolasHug
Copy link
Contributor

It seems from https://github.com/pytorch/torchcodec/actions/runs/17402821403/job/49401670135?pr=831 that test-infra stopped support 12.9 and is generating jobs for 13.0 instead, so we need to change our jobs to reflect that.

@meta-cla meta-cla bot added the CLA Signed This label is managed by the Meta Open Source bot. label Sep 2, 2025
# PR.
python-version: ['3.10']
cuda-version: ['12.6']
cuda-version: ['12.8']
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Drive-by, 12.6 is still supported but the "default" CUDA version supported by pytorch is currently 12.8

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Its possible that the github runner linux.4xlarge.nvidia.gpu does not have CUDA 12.8, and thats why this job was erroring. I've reverted it since its not related to this change, but we could leave a TODO to update the CUDA version if its something we need to change eventually.

@NicolasHug NicolasHug marked this pull request as draft September 2, 2025 14:28
if sys.platform == "linux":
if args[0].device.type == "cuda":
atol = 2
atol = 3
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think we should be able to preserve the previous stricter with CUDA < 13 with something like:

Suggested change
atol = 3
atol = 3 if cuda_version_used_for_building_torch() >= (13, 0) else 2

@NicolasHug
Copy link
Contributor Author

Thanks @Dan-Flores for the fixes! LGTM, let's close this one so you can open a new one

@NicolasHug NicolasHug closed this Sep 30, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
CLA Signed This label is managed by the Meta Open Source bot.
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants