Add PTX code for (highest) CUDA compute capability in PyTorch & torchvision - #4129
Open
Flamefire wants to merge 2 commits into
Open
Add PTX code for (highest) CUDA compute capability in PyTorch & torchvision#4129Flamefire wants to merge 2 commits into
Flamefire wants to merge 2 commits into
Conversation
Flamefire
force-pushed
the
pytorch-ptx-code
branch
from
May 6, 2026 07:23
687e6c2 to
060e0a5
Compare
…vision E.g. "5.0+PTX" makes PyTorch add PTX code. We can simply add it to the list in `TORCH_CUDA_ARCH_LIST` to add PTX code for the last architecture.
Flamefire
force-pushed
the
pytorch-ptx-code
branch
from
July 15, 2026 13:49
060e0a5 to
8a6b5b5
Compare
Makes it easier to change for (interactive) tests.
Flamefire
force-pushed
the
pytorch-ptx-code
branch
from
July 16, 2026 12:38
8a6b5b5 to
9ef672d
Compare
boegel
requested changes
Jul 29, 2026
| ', '.join(cuda_arch_list.split(';'))) | ||
| if cuda_arch_list: | ||
| self.log.info('Also creating PTX code for architecture ' + cuda_arch_list.split(';')[-1]) | ||
| cuda_arch_list += '+PTX' |
Member
There was a problem hiding this comment.
As currently proposed, we'll start doing this for every PyTorch version...
Maybe we should only do this for (very) recent PyTorch versions (say 2.9+), along with a way to explicitly opt-in to doing this?
Contributor
Author
There was a problem hiding this comment.
IIRC we new cuda sanity checks check that the highest CC has PTX code and fails otherwise. So to me this seems like a bugfix
If you want I can run some builds with e.g. oldest and newest 1.x and 2.x version we have or select some randomly. I'd expect it to either work everywhere or fail everywhere as the biggest "risk" I see is if some PyTorch version didn't support it yet although it seems it's been always the case.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
E.g. "5.0+PTX" makes PyTorch add PTX code.
We can simply add it to the list in
TORCH_CUDA_ARCH_LISTto add PTX code for the last architecture.Uses parts of #4092. So with easybuilders/easybuild-framework#5144 it will add PTX code for the highest arch