mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #324] Why kNVLReceiver warps in Internode Normal Dispatch use TMA #226
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#226
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 @ZhiyiHu1999 on GitHub (Jul 22, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/324
Hello! In my comprehension, TMA is an asynchronous copy and can be used to decouple NVLink data transmission to achieve further communication-computation overlap. In current implementation, internode normal dispatch uses TMA in
kRDMAAndNVLForwarderwarps andkNVLReceiverwarps, and internode normal combine uses TMA in 'kNVLSender' warps.However, data transmission in
kNVLReceiverwarps happen on the same GPU and I think using TMA (global mem -> shared mem -> global mem) introduces more cost than the previous warp copy (global mem -> global mem). So, is there any other purpose for the TMA use ofkReceiverwarps in normal mode internode dispatch.Thanks a lot!