mirror of
https://github.com/deepseek-ai/DeepGEMM.git
synced 2026-11-05 00:47:26 +00:00
[GH-ISSUE #199] The comparison between cublaslt and deepgemm in test_fp8.py is unaligned #114
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepGEMM#114
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @Insideyyy on GitHub (Sep 26, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/199
It's great that PR #198 added cublas implementation to test_fp8.py, so we can directly observe the performance differences.
That said, I noticed the cublas implementation uses tensorwise scaling(cublas default behavior) instead of blockwise scaling, which makes the comparison with deepgemm less fair and potentially misleading.
Would it be possible to align the scaling strategy? I suppose cublas currently has supported the blockwise scaling strategy same with deepgemm: https://docs.nvidia.com/cuda/cublas/#element-1d-and-128x128-2d-block-scaling-for-fp8-data-types
@zheanxu commented on GitHub (Apr 27, 2026):
Thanks for the detailed observation! This was intentional: we chose tensorwise scaling to keep the cuBLAS baseline simple and use it as a "no block scaling overhead" reference point for comparison.