[GH-ISSUE #175] Question: it seems that DeepSeek V3.1 is faster than DeepSeek V3(0324 version) when decoding under SGLang and H800. #98

Closed
opened 2026-10-26 21:12:31 +00:00 by bsakaguchi · 2 comments
Owner

Originally created by @Huixxi on GitHub (Aug 28, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/175

I don't know whether V3.1 is really faster than V3, but if it has something to do with its new UE8M0 and DeepGEMM is support this type of FP8 natively? Thanks.

Originally created by @Huixxi on GitHub (Aug 28, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/175 I don't know whether V3.1 is really faster than V3, but if it has something to do with its new UE8M0 and DeepGEMM is support this type of FP8 natively? Thanks.
Author
Owner

@LyricZhao commented on GitHub (Aug 28, 2025):

On the same device, UE8M0 and FP32 SF offer the same level of performance. I guess the reason is that V3.1's output length is shorter than V3 averagely.

<!-- gh-comment-id:3232591897 --> @LyricZhao commented on GitHub (Aug 28, 2025): On the same device, UE8M0 and FP32 SF offer the same level of performance. I guess the reason is that V3.1's output length is shorter than V3 averagely.
Author
Owner

@Huixxi commented on GitHub (Aug 28, 2025):

Ok, thanks a lot.

<!-- gh-comment-id:3232637185 --> @Huixxi commented on GitHub (Aug 28, 2025): Ok, thanks a lot.
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#98
No description provided.