Repository navigation
[MSLK] Honor CUDA architecture selection for source builds - #1334
Merged
Merged
Conversation
xuzhao9
approved these changes
Nov 4, 2026
xuzhao9
left a comment
xuzhao9
left a comment
Contributor
There was a problem hiding this comment.
Thanks for addressing this!
Just FYI, our CI pipeline are now using the nightly mslk build, so we are no longer using self-built MSLK in our testing environment.
0z5a
had a problem deploying
to
docker-s3-upload
November 4, 2026 16:58 — with 
GitHub Actions
Failure
GitHub Actions
Failure
0z5a
marked this pull request as ready for review
November 4, 2026 17:06
This branch was previously deployed
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The MSLK source installer hard-codes a CUDA architecture list even when
TORCH_CUDA_ARCH_LISTis set. Honor a nonempty user selection and keep the existing list as the default, allowing source builds such as B300's10.3atarget.The nightly Docker workflow also uses a pull request's branch name while cloning the upstream repository. For PRs from forks this fails before the build starts. Pass the PR head repository to the Dockerfile along with its branch; scheduled and manually dispatched builds still default to the upstream repository.
Fixes #1333.
Testing
10.3a, and a mixed9.0a;10.3alist. Both explicit-selection cases failed before the installer change.TORCH_CUDA_ARCH_LIST=10.3aagainst a previously compiled pinned MSLK source tree and confirmed installation completed. The cached build used CUDA 13.3.1; CUDA 13.0 was present for the install-time toolkit version check.ufmt checkpassed for the installer and its regression test. The Docker workflow YAML and build-argument wiring were checked locally, and the fork branch was cloned at the previous PR head. A full Docker build was not run locally because the Docker daemon was unavailable.Prebuilt wheel coverage and benchmark gating are separate from source-build target selection.