Repository navigation
Conversation
denera
force-pushed
the
pytorch/fp8bs-grouped-gemm-hopper-fix
branch
from
October 21, 2026 16:27
0d14e26 to
49df5dd
Compare
Collaborator
Author
|
/te-ci pytorch |
Contributor
 [Low impact] The PR appears safe to merge; no new actionable issue remains. SummaryChanges the FP8 block-scaling grouped GEMM tests to compare against independently quantized split GEMMs instead of FP32 multiplication.
Reviews (2) · Last reviewed commit: "[PyTorch] Test grouped FP8 block-scaling..." · Reviewed by Greptile |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
…r path MXFP8 GEMM on Blackwell accumulates at higher precision than FP8 block-scaling GEMM on Hopper. The grouped FP8 block-scaling GEMM tests compared against an FP32 reference, so they passed with the MXFP8 emulation on Blackwell but failed on Hopper. Compare against the legacy split-tensor grouped GEMM instead, consistent with the existing MXFP8 grouped GEMM tests. Also remove the FP8 block-scaling zero-size-expert and mixed-operand tests, which no other recipe has. Signed-off-by: Alp Dener <adener@nvidia.com>
denera
force-pushed
the
pytorch/fp8bs-grouped-gemm-hopper-fix
branch
from
October 21, 2026 17:17
49df5dd to
2717644
Compare
Collaborator
Author
|
/te-ci pytorch |
This branch has not been 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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Description
FP8BS grouped GEMM tests were being compared against a reference with FP32 accumulation. FP8BS on Blackwell via MXFP8 emulation passed these tests because MXFP8 grouped GEMM on Blackwell accumulates at a higher precision than native FP8BS grouped GEMM on Hopper.
This PR changes the reference in the test to the unfused split-grouped GEMM — consistent with how native MXFP8 grouped GEMM is tested.
Redundant zero-token expert and mixed precision rejection tests were removed because these are already covered under different test cases.
Type of change
Checklist: