[GH-ISSUE #575] deep_ep::intranode::cached_notify_combine<8> takes very long #354

New issue

Open

opened 2026-10-26 09:59:00 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:59:00 +00:00

Owner

Originally created by @goldhuang on GitHub (Jan 27, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/575

Hi team,

I'm doing intronode EP with deepep. But cached_notify_combine() could takes 13-23ms, while the grouped gemm only taks less than 3ms. Do you have any insight about potential reasons?
deep_ep::intranode::combine() takes 7ms or less.
deep_ep::intranode::dispatch() takes around 5ms.

Thanks a lot!

Originally created by @goldhuang on GitHub (Jan 27, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/575 Hi team, I'm doing intronode EP with deepep. But cached_notify_combine() could takes 13-23ms, while the grouped gemm only taks less than 3ms. Do you have any insight about potential reasons? deep_ep::intranode::combine() takes 7ms or less. deep_ep::intranode::dispatch() takes around 5ms. Thanks a lot!

bsakaguchi commented

2026-10-26 09:59:02 +00:00

Author

Owner

@baichenhan commented on GitHub (Feb 2, 2026):

The load may be unbalanced.

@baichenhan commented on GitHub (Feb 2, 2026): The load may be unbalanced.

bsakaguchi referenced this issue

2026-10-27 21:03:44 +00:00

[PR #354] [CLOSED] fix #532