mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #513] [Question] Modifications to achieve concurrent dispatch and combine in low latency mode #332
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepEP#332
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @tianr22 on GitHub (Dec 5, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/513
Hello!
Currently I am trying to use DeepEP to achieve three batch overlap, in which I need to issue the dispatch_send phase of batch 1 and combine_send phase of batch 0 concurrently.
I understand current implementation can not achieve this since we we only have 2 lowlatency buffer and will clean the metadata buffer(such as rdma_recv_flag_buffer) of the next buffer when we use one buffer to do actual send and recv. When dispatch_send and combine_send issue together, the later one will clean the metabuffer of former, which means the former operation may never finish recv(hang at checking the buffer value).
i wonder why u fuse the clean next buffer operation into dispatch/combine, instaed of a seperate kernel( for the consideration of performance?)
Can u give me some suggestions on how to achieve the concurrent dispatch and combine in low latency mode?
I thought one potential way was to clean the current metabuffer when recv kernel finishes. The kernel does nothing to the next buffer. But will there be some bug?
Thanks!