xformers v0.0.31 and rebuild for pytorch 2.7#50
Conversation
|
@conda-forge-admin, please rerender |
|
Hi! This is the friendly automated conda-forge-linting service. I just wanted to let you know that I linted all conda-recipes in your PR ( I do have some suggestions for making it better though... For recipe/recipe.yaml:
This message was generated by GitHub Actions workflow run https://github.com/conda-forge/conda-forge-webservices/actions/runs/16443451998. Examine the logs at this URL for more detail. |
…nda-forge-pinning 2025.06.04.10.18.16
|
@conda-forge-admin, please rerender |
…nda-forge-pinning 2025.06.04.10.18.16
|
@conda-forge-admin, please rerender |
…nda-forge-pinning 2025.06.04.10.18.16
jeongseok-meta
changed the title
|
@conda-forge-admin, please rerender |
…nda-forge-pinning 2025.06.25.17.20.38
export CUTLASS_NVCC_ARCHS="80 86 89"
|
I had to restrict CUDA archs to 80/86/89 because compiling for sm90 makes xformers v0.0.31 crash (NVCC exit 255). It might not be the ideal fix, but the build otherwise fails as: |
jeongseok-meta
changed the title
| export TORCH_CUDA_ARCH_LIST="5.3;6.0;6.1;7.0;7.5;8.0;8.6;8.9;9.0+PTX" | ||
| if [[ ${cuda_compiler_version} == 12.6 ]]; then | ||
| export TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9" | ||
| export CUTLASS_NVCC_ARCHS="80 86 89" |
There was a problem hiding this comment.
I don't think this is acceptable TBH. We need to keep support for more architectures, we shouldn't drop 9.0, and we also shouldn't drop the +PTX. Where do CUTLASS_NVCC_ARCHS come into play?
There was a problem hiding this comment.
FYI: It builds with export TORCH_CUDA_ARCH_LIST="5.3;6.0;6.1;7.0;7.5;8.0;8.6;8.9", without defining CUTLASS_NVCC_ARCHS.
Not sure if we need CUDA 12.8+ for sm90 builds, but I also got an error once I added 9.0+PTX back in.
There was a problem hiding this comment.
Yes, it appears that CUTLASS_NVCC_ARCHS is not necessary. This was the only green path I could find, but I welcome any suggestions or fixes as long as they work.
There was a problem hiding this comment.
I'm trying a build with CUDA 12.8 and adding 9.0+PTX back in, but assuming that nothing crashes, we'll need to wait ~5 hours before we see the results.
There was a problem hiding this comment.
@shermansiu Thanks for taking a look at this. Feel free to directly modify this PR, by the way.
Oh, you are not a maintainer. Feel free to create another PR if you prefer then.
There was a problem hiding this comment.
Also, I think we can add 5.3;6.0;6.1;7.0;7.5; back into TORCH_CUDA_ARCH_LIST?
There was a problem hiding this comment.
I could, but I think we're almost done? I think we can drop 9.0+PTX for the CUDA 12.6 build (this PR) and add it back in once we add the CUDA 12.8 one, in a different PR (assuming it compiles).
There was a problem hiding this comment.
Update: Using CUDA 12.8 did not work for 9.0+PTX... For debugging purposes, I'm just targeting that arch so I can figure out why it's throwing an error
|
@conda-forge-admin, please rerender |
…nda-forge-pinning 2025.07.21.12.24.54
This reverts commit dc711be.
…nda-forge-pinning 2025.07.21.12.24.54
h-vetinari
left a comment
h-vetinari
left a comment
There was a problem hiding this comment.
Thanks for the work here @jeongseok-meta!
|
@h-vetinari Thank you for completing and merging this PR! Also, many thanks to @shermansiu for your valuable input! |
|
Good news: sm_90 works again when compiled with CUDA 12.9. It's been re-added in #52, along with new architectures sm_100 and sm_120. The CUDA builds now need ~9 hours, but OK, that's what we got the server for. 🙃 |
v0.0.30 is present in the git history, but was never published, since #50 was merged after bumping to v0.0.31. Squash of all the commits on main branch since then --------- Co-authored-by: Jeongseok Lee <jeongseok@meta.com> Co-authored-by: H. Vetinari <h.vetinari@gmx.com>
4 participants
Checklist
0(if the version changed)conda-smithy(Use the phrase@conda-forge-admin, please rerenderin a comment in this PR for automated rerendering)Closes #43
Closes #44
Closes #47
Closes #49
Closes #51