deepseek-ai/DeepEP

Fork 0

mirror of https://github.com/deepseek-ai/DeepEP.git synced 2026-11-05 00:47:32 +00:00

[GH-ISSUE #197] How much performance can TMA optimization gain on intra-node normal kernel? #160

New issue

Closed

opened 2026-10-26 09:55:24 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-26 09:55:24 +00:00

Owner

Originally created by @wzc-wuzhicheng on GitHub (Jun 9, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/197

I notice that you add TMA optimization on intra-node normal kernels at pr https://github.com/deepseek-ai/DeepEP/pull/191.

I tested it in my environment by python tests/test_intranode.py but found no significant performance improvement. I'm not sure if I missed some steps, can you provide any performance data for reference? Thank you very much.

What's more, can TMA optimization be applied to internode normal kernel and How much potential performance improvement will it gain ?

Originally created by @wzc-wuzhicheng on GitHub (Jun 9, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/197 I notice that you add TMA optimization on intra-node normal kernels at pr https://github.com/deepseek-ai/DeepEP/pull/191. I tested it in my environment by `python tests/test_intranode.py` but found no significant performance improvement. I'm not sure if I missed some steps, can you provide any performance data for reference? Thank you very much. What's more, can TMA optimization be applied to internode normal kernel and How much potential performance improvement will it gain ?

bsakaguchi closed this issue

2026-10-26 09:55:24 +00:00

bsakaguchi commented

2026-10-26 09:55:25 +00:00

Author

Owner

@alpha-baby commented on GitHub (Jun 10, 2025):

I have the same question, can you help me to answer it? Thank you very much. @LyricZhao

@alpha-baby commented on GitHub (Jun 10, 2025): I have the same question, can you help me to answer it? Thank you very much. @LyricZhao

bsakaguchi commented

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

Author

Owner

@LyricZhao commented on GitHub (Jun 10, 2025):

I tested it in my environment by python tests/test_intranode.py but found no significant performance improvement. I'm not sure if I missed some steps, can you provide any performance data for reference? Thank you very much.

It is as expected. See https://github.com/deepseek-ai/DeepEP/issues/184#issuecomment-2955323318.

What's more, can TMA optimization be applied to internode normal kernel and How much potential performance improvement will it gain?

Yes, and we are still working on this. The exact number will be released later, maybe in 2 weeks.

@LyricZhao commented on GitHub (Jun 10, 2025): > I tested it in my environment by python tests/test_intranode.py but found no significant performance improvement. I'm not sure if I missed some steps, can you provide any performance data for reference? Thank you very much. It is as expected. See https://github.com/deepseek-ai/DeepEP/issues/184#issuecomment-2955323318. > What's more, can TMA optimization be applied to internode normal kernel and How much potential performance improvement will it gain? Yes, and we are still working on this. The exact number will be released later, maybe in 2 weeks.

bsakaguchi commented

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

Author

Owner

@wzc-wuzhicheng commented on GitHub (Jun 11, 2025):

I tested it in my environment by python tests/test_intranode.py but found no significant performance improvement. I'm not sure if I missed some steps, can you provide any performance data for reference? Thank you very much.

It is as expected. See #184 (comment).

What's more, can TMA optimization be applied to internode normal kernel and How much potential performance improvement will it gain?

Yes, and we are still working on this. The exact number will be released later, maybe in 2 weeks.

OK, Thanks for your reply!

@wzc-wuzhicheng commented on GitHub (Jun 11, 2025): > > I tested it in my environment by python tests/test_intranode.py but found no significant performance improvement. I'm not sure if I missed some steps, can you provide any performance data for reference? Thank you very much. > > It is as expected. See [#184 (comment)](https://github.com/deepseek-ai/DeepEP/issues/184#issuecomment-2955323318). > > > What's more, can TMA optimization be applied to internode normal kernel and How much potential performance improvement will it gain? > > Yes, and we are still working on this. The exact number will be released later, maybe in 2 weeks. OK, Thanks for your reply!

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/DeepEP#160

No description provided.

Rows
Columns