[GH-ISSUE #259] Question Regarding Low-Latency Buffer Layout #195

New issue

Closed

opened 2026-10-26 09:56:22 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:56:22 +00:00

Owner

Originally created by @ZhiyiHu1999 on GitHub (Jun 26, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/259

Hello! I want to ask why num_experts is not multiplied to calculate dispatch_send_buffer_bytes. I think for the worst case that every expert may have the chance to be sent all dispatched tokens.

github.com/deepseek-ai/DeepEP@bd429ffefc/csrc/config.hpp (L144-L148)

Thanks a lot!

Originally created by @ZhiyiHu1999 on GitHub (Jun 26, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/259 Hello! I want to ask why `num_experts` is not multiplied to calculate `dispatch_send_buffer_bytes`. I think for the worst case that every expert may have the chance to be sent all dispatched tokens. https://github.com/deepseek-ai/DeepEP/blob/bd429ffefc50cad10b5b17d63eed47c0ab8db72a/csrc/config.hpp#L144-L148 Thanks a lot!

bsakaguchi closed this issue

2026-10-26 09:56:22 +00:00

bsakaguchi commented

2026-10-26 09:56:23 +00:00

Author

Owner

@ZhiyiHu1999 commented on GitHub (Jun 26, 2025):

I found out. Thanks a lot!

@ZhiyiHu1999 commented on GitHub (Jun 26, 2025): I found out. Thanks a lot!

bsakaguchi referenced this issue

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

[PR #195] [MERGED] Add low-latency kernel PCIe usage flag #458