[GH-ISSUE #178] Question about "SM-free normal kernels" in Roadmap #153

New issue

Closed

opened 2026-10-26 09:55:17 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 09:55:17 +00:00

Owner

Originally created by @polarstormx on GitHub (May 28, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/178

Hello DeepEP Team,

I noticed "SM-free normal kernels" mentioned in your Roadmap.

I'd like to clarify:

Does “SM-free” mean the kernels will avoid SM usage entirely, or does it simply indicate a large reduction in SM load (e.g., saving nvlink forworder by using copy engine)?
Is there a planned release date or timeline for making this feature available?

Thanks!

Originally created by @polarstormx on GitHub (May 28, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/178 Hello DeepEP Team, I noticed "SM-free normal kernels" mentioned in your Roadmap. I'd like to clarify: 1. Does “SM-free” mean the kernels will avoid SM usage entirely, or does it simply indicate a large reduction in SM load (e.g., saving nvlink forworder by using copy engine)? 2. Is there a planned release date or timeline for making this feature available? Thanks!

bsakaguchi closed this issue

2026-10-26 09:55:17 +00:00

bsakaguchi commented

2026-10-26 09:55:17 +00:00

Author

Owner

@LyricZhao commented on GitHub (May 29, 2025):

As combine must be done by GPU, so maybe it is a large reduction in SM load for combine. But for dispatch, we are still figuring out whether we can avoid SM entirely. Or another solution is to use warps to overlap, e.g. within a SM, 12 warps do GEMM, 4 warps do communication (but it requires some tricks to make registers/shared memory perfectly allocated, I am not sure whether it could work).
Sorry, no about this. Maybe by the end of this year? We are working on something else recently.

@LyricZhao commented on GitHub (May 29, 2025): 1. As combine must be done by GPU, so maybe it is a large reduction in SM load for combine. But for dispatch, we are still figuring out whether we can avoid SM entirely. Or another solution is to use warps to overlap, e.g. within a SM, 12 warps do GEMM, 4 warps do communication (but it requires some tricks to make registers/shared memory perfectly allocated, I am not sure whether it could work). 2. Sorry, no about this. Maybe by the end of this year? We are working on something else recently.

bsakaguchi commented

2026-10-26 09:55:18 +00:00

Author

Owner

@polarstormx commented on GitHub (May 29, 2025):

Thanks for your reply!

@polarstormx commented on GitHub (May 29, 2025): Thanks for your reply!

bsakaguchi referenced this issue

2026-10-26 10:00:49 +00:00

[PR #153] [MERGED] Shuffling the starting index of target rank for different ranks and channels #444