[GH-ISSUE #296] [Question] How to autotune DeepGemm FP8 to maximize this baseline performance? #148

Closed
opened 2026-10-26 21:13:43 +00:00 by bsakaguchi · 1 comment
Owner

Originally created by @ghostplant on GitHub (Mar 25, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/296

To maximize the performance of DeepGemm, do I need to call some pre-configuration / autotuning interface to maximize this performance for specific shape below?

# auto-tune needed or not?
deep_gemm.fp8_gemm_nt((a_fp8, a_s), (b_fp8, out1)
Originally created by @ghostplant on GitHub (Mar 25, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/296 To maximize the performance of DeepGemm, do I need to call some pre-configuration / autotuning interface to maximize this performance **for specific shape** below? ``` # auto-tune needed or not? deep_gemm.fp8_gemm_nt((a_fp8, a_s), (b_fp8, out1) ```
Author
Owner

@zheanxu commented on GitHub (Apr 17, 2026):

The FP8 GEMM automatically selects a good config based on built-in heuristics, so no explicit autotuning interface is required. If you want to achieve the absolute best performance for your specific shape, you can manually adjust the heuristics.

<!-- gh-comment-id:4264912850 --> @zheanxu commented on GitHub (Apr 17, 2026): The FP8 GEMM automatically selects a good config based on built-in heuristics, so no explicit autotuning interface is required. If you want to achieve the absolute best performance for your specific shape, you can manually adjust the heuristics.
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#148
No description provided.