Skip to content

fix(plugin): expose backend ops required by TP comm overlap - #95

Merged
zhaoyinglia merged 4 commits into
flagos-ai:mainfrom
lihongyang1990:fix/tp-comm-overlap-backend-ops
Aug 20, 2026
Merged

fix(plugin): expose backend ops required by TP comm overlap#95
zhaoyinglia merged 4 commits into
flagos-ai:mainfrom
lihongyang1990:fix/tp-comm-overlap-backend-ops

Conversation

@lihongyang1990

Copy link
Copy Markdown

Summary

  • register device_supports_multicast, ubuf_built_with_mpi, and get_stream_priority_range for the CUDA and Hygon plugin backends
  • forward the runtime queries to each native backend extension
  • convert CommOverlapType to the native backend enum before creating CommOverlapP2P

Problem

When tp_comm_overlap: true is enabled, user-buffer initialization calls these runtime helpers through the TE-FL plugin interface. The CUDA and Hygon implementations were not exposed in the operator registry, which caused missing-function errors before TP communication overlap could initialize.

Validation

  • compiled all four changed Python modules
  • git diff --check

@zhaoyinglia
zhaoyinglia merged commit 71cd511 into flagos-ai:main Aug 20, 2026
55 of 61 checks passed
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.

2 participants