mirror of
https://github.com/deepseek-ai/DeepEP.git
synced 2026-11-05 00:47:32 +00:00
[GH-ISSUE #436] notify_dispatch and cached_notify in normal mode. #290
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#290
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 @MaoZiming on GitHub (Sep 27, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/436
Hi, I am wondering why compared to low latency mode, normal mode needs additional synchronization primitives, such as
nvshmem_sync_with_same_gpu_idxandnvshmemi_ibgda_quietfor flushing, whereas low latency mode does not have that.Can normal mode follow low latency mode in not using these primitives?
@a1372422617 commented on GitHub (Sep 28, 2025):
Because the normal mode needs to ensure that the preceding RDMA operations are written to disk, it uses a barrier to guarantee that all GPUs have reached the synchronization point before clearing the flag to zero
@MaoZiming commented on GitHub (Sep 29, 2025):
Thanks for the answer. What do you mean by "written to disk"? I am wondering why low latency mode does not need these primitives
@sphish commented on GitHub (Oct 10, 2025):
The low-latency mode uses a double-buffer design specifically to avoid relying on these synchronization primitives. In contrast, the normal kernel does not use double buffering — this is a trade-off between GPU memory usage and latency.