[GH-ISSUE #82] What is the internally and carefully optimized implementation based on CUTLASS 3.6? #47

Closed
opened 2026-10-26 21:11:11 +00:00 by bsakaguchi · 1 comment
Owner

Originally created by @lu229 on GitHub (Apr 9, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/82

I test deepGEMM on H20, and find it is faster than cutlass fp8 groupwise GEMM, but is slower than cutlass fp8 GEMM. I want to know What is the internally and carefully optimized implementation based on CUTLASS 3.6? Is it an internal version(support fp8 fine-grained GEMM) of DeepSeek?

Originally created by @lu229 on GitHub (Apr 9, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/82 I test deepGEMM on H20, and find it is faster than cutlass fp8 groupwise GEMM, but is slower than cutlass fp8 GEMM. I want to know What is the internally and carefully optimized implementation based on CUTLASS 3.6? Is it an internal version(support fp8 fine-grained GEMM) of DeepSeek?
Author
Owner

@soundOfDestiny commented on GitHub (Apr 10, 2025):

yeah we need fine-grained scaling

<!-- gh-comment-id:2791277841 --> @soundOfDestiny commented on GitHub (Apr 10, 2025): yeah we need fine-grained scaling
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepGEMM#47
No description provided.