Skip to content

fix(dp): wait forward before gather_req - #1506

Open
WANDY666 wants to merge 1 commit into
mainfrom
fix/dp-collective-overlap-ordering
Open

fix(dp): wait forward before gather_req#1506
WANDY666 wants to merge 1 commit into
mainfrom
fix/dp-collective-overlap-ordering

Conversation

@WANDY666

Copy link
Copy Markdown
Contributor

Summary

  • Serialize the next DP collective behind GPU work already queued on the previous overlap stream.
  • Prevent NCCL all-gathers from overlapping outstanding DeepEP/MTP kernels.
  • Preserve request-preparation overlap without adding a host-side synchronization.

Root cause

With overlap enabled, the previous infer thread can release forward before its GPU work finishes. The next infer-loop iteration may then enqueue DP all-gathers on the current stream while the prior DeepEP/MTP kernels are still running. Adding an explicit stream dependency restores the required ordering.

This PR isolates the fix from d5b04bf onto the latest main; the resulting commit is b98f6802.

Testing

  • Focused mocked DP overlap-ordering regression test passed in the H100 LightLLM environment.
  • Changed module passes Python byte-compilation.
  • git diff --check passes.

The next infer thread may launch DP all-gathers before the previous
overlap-stream forward finishes, causing NCCL to overlap with DeepEP/MTP
kernels. Make the current stream wait before starting DP collectives
while preserving overlap for request preparation.
@WANDY666 WANDY666 changed the title fix(dp): serialize DP collectives after overlap forward fix(dp): wait forward before gather_req Aug 26, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant