[GH-ISSUE #669] [Performance] Question about performance of Deepep v2, the role of the prefer-overlap-with-compute parameter #394

Open
opened 2026-10-26 10:00:01 +00:00 by bsakaguchi · 3 comments
Owner

Originally created by @shenhongqian on GitHub (Jun 22, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/669

I would like to ask, what are the execution commands for the performance of the two machines shown in the figure below, and do they need to include --prefer-overlap-with-compute 1 ?

Image

I am not aligned with the official performance data in my local environment, and the following are the test results.
[1]python tests/elastic/test_ep.py --num-processes 8 --allow-hybrid-mode 1 --num-sms 12 --reuse-elastic-buffer --num-topk 8 --num-tokens 8192 --prefer-overlap-with-compute 1 --num-qps 65

Image

[2]python tests/elastic/test_ep.py --num-processes 8 --allow-hybrid-mode 1 --num-sms 12 --reuse-elastic-buffer --num-topk 8 --num-tokens 8192 --prefer-overlap-with-compute 0 --num-qps 65

Image
Originally created by @shenhongqian on GitHub (Jun 22, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/669 I would like to ask, what are the execution commands for the performance of the two machines shown in the figure below, and do they need to include --prefer-overlap-with-compute 1 ? <img width="1695" height="1245" alt="Image" src="https://github.com/user-attachments/assets/307983c1-2c3e-42cc-ad7b-bf928a0bb814" /> I am not aligned with the official performance data in my local environment, and the following are the test results. [1]python tests/elastic/test_ep.py --num-processes 8 --allow-hybrid-mode 1 --num-sms 12 --reuse-elastic-buffer --num-topk 8 --num-tokens 8192 -**-prefer-overlap-with-compute 1** --num-qps 65 <img width="2151" height="1509" alt="Image" src="https://github.com/user-attachments/assets/10e3af86-4ccf-48f9-aff2-25dfa5e47ab8" /> [2]python tests/elastic/test_ep.py --num-processes 8 --allow-hybrid-mode 1 --num-sms 12 --reuse-elastic-buffer --num-topk 8 --num-tokens 8192 -**-prefer-overlap-with-compute 0** --num-qps 65 <img width="2076" height="1497" alt="Image" src="https://github.com/user-attachments/assets/e9babbeb-96d8-49ab-8980-5cede77e5e5a" />
Author
Owner

@shenhongqian commented on GitHub (Jun 22, 2026):

Can you help me take a look at this issue? Thank you @sphish @LyricZhao

<!-- gh-comment-id:4764637992 --> @shenhongqian commented on GitHub (Jun 22, 2026): Can you help me take a look at this issue? Thank you @sphish @LyricZhao
Author
Owner

@alpha-baby commented on GitHub (Jun 23, 2026):

maybe you could try SMs == 16 , and will Improve the performance of combine

<!-- gh-comment-id:4775463608 --> @alpha-baby commented on GitHub (Jun 23, 2026): maybe you could try SMs == 16 , and will Improve the performance of combine
Author
Owner

@kzlxd commented on GitHub (Jul 27, 2026):

@shenhongqian --prefer-overlap-with-compute 1 is necessary

<!-- gh-comment-id:5089461652 --> @kzlxd commented on GitHub (Jul 27, 2026): @shenhongqian --prefer-overlap-with-compute 1 is necessary
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/DeepEP#394
No description provided.