[RL] Fix weight update for mxfp8 flashinfer_cutlass gemm backend#22484
Merged
b8zhong merged 3 commits intosgl-project:mainfrom Apr 12, 2026
Merged
[RL] Fix weight update for mxfp8 flashinfer_cutlass gemm backend#22484b8zhong merged 3 commits intosgl-project:mainfrom
b8zhong merged 3 commits intosgl-project:mainfrom
Conversation
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Contributor
Author
|
Hi @b8zhong , could you take a look at the tiny reverting fix? Thank you! |
355f101 to
2f03d08
Compare
b8zhong
approved these changes
Apr 11, 2026
Collaborator
|
/tag-and-rerun-ci |
Contributor
Author
|
/rerun-failed-ci |
2 similar comments
Contributor
Author
|
/rerun-failed-ci |
Contributor
Author
|
/rerun-failed-ci |
30 tasks
pyc96
pushed a commit
to pyc96/sglang
that referenced
this pull request
Apr 14, 2026
yhyang201
pushed a commit
to yhyang201/sglang
that referenced
this pull request
Apr 22, 2026
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
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
@HumansAnd
#21576 refactors mxfp8 scaling factor swizzling to in-place style. However, on the
flashinfer_cutlassmxfp8 code path,block_scale_interleavemay pad the scales, violating the shape contract for weight update. Therefore, we revert to the previous original & swizzled dual buffer approach.For the full mxfp8 DeepSeek 671B model, all the duplicate
ue8m0accounts for less than 1gb, thus the overhead is negligible.In the future we should rely on a
restore_weights_before_loadingapi which is still under development.Modifications
flashinfer_cutlassmxfp8 gemm to style prior [FlashInver v0.6.7] Integrate flashinfer_trtllm mxfp8 gemm #21576Accuracy Tests
Speed Tests and Profiling
Checklist
Review and Merge Process
/tag-and-rerun-ci,/tag-run-ci-label,/rerun-failed-ci