mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #335] [Question] Use only RDMA non low-latency #233
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#233
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 @ybenvidia on GitHub (Jul 28, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/335
Is there a way to force dispatch to use only RDMA even for intranode in non low-latency mode ?
If yes, can. I do it in runtime (using for example an env variable)
Thanks
@sphish commented on GitHub (Jul 29, 2025):
It's not supported.
@ybenvidia commented on GitHub (Jul 29, 2025):
What if I put NUM_MAX_NVL_PEERS = 1 ?
@sphish commented on GitHub (Jul 30, 2025):
You can give it a try, but our code has a limit on the
num_rdma_ranks. So after you setNUM_MAX_NVL_PEERS = 1, the applicable cases will be quite limited.@ybenvidia commented on GitHub (Aug 11, 2025):
Another question, from what I saw in the code we can only use your code with 20 nodes max. Is it correct ? What if we want to use more than 20 nodes ? The performance will not be good ?
@MengYu10151 commented on GitHub (Nov 14, 2025):
@ybenvidia hi, hope its not too late. We implement a kernel to support RDMA only in normal mode, u can refer to https://github.com/deepseek-ai/DeepEP/pull/375, already merged into hybrid-ep branch :)