fix(weight-sync): compose engine waves, credits and timeline - #2362
Open
0z5a wants to merge 4 commits into
Open
Conversation
Keep logical credit, transport, staging, and consumer lifetimes distinct; propagate colocated load acknowledgement across each existing TP group; poison failed versions; and publish only after all consumers resume. Add fixed-cardinality metrics, failure coverage, and opt-in 2/4-rank NCCL validation. Assisted-by: OpenAI Codex Signed-off-by: jiachun <2609024259@qq.com>
Integrate the local wave/callsite stack through 175ad4c with the credit stack through e9fd2ee. Preserve logical bucket admission across waves, fence transport completion, propagate load failure before the next wave, retain failed-wave resources, and keep version publication after consumer resume. Validate both production updater variants with logical 2/4-engine tests and real Gloo credit transport. Assisted-by: OpenAI Codex Signed-off-by: jiachun <2609024259@qq.com>
…ycles Integrate the timeline stack through 100470c on top of wave/credit composition. Preserve API return separately from transport completion, close NVTX ranges before subsequent submissions, retain wave identity, trace consumer acknowledgements and actual reuse/publication boundaries, and record failures without success markers. Cover trace-on/off production updater contracts; GPU/Ray/SGLang end-to-end confirmation remains unvalidated. Assisted-by: OpenAI Codex Signed-off-by: jiachun <2609024259@qq.com>
Author
|
Hi @zhuzilin Would you mind having a review? |
0z5a
marked this pull request as ready for review
September 7, 2026 09:35
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Compose the engine-wave, bucket-credit and trainer-timeline proposals in #2351, #2352 and #2353. This is a Draft integration follow-up, not an additional competing scheduler. The new integration commits are 6ce21e4 and 58d5048; GitHub's main-based aggregate diff also includes their prerequisite stacks.
Validation
Recorded on this exact integration revision before publication:
The CPU engine fixtures and fake-CUDA event tests are not real SGLang loads. Full trainer-to-conversion-to-Ray/SGLang GPU execution, full-model numerical checks and independent performance confirmation remain outstanding. No new GPU run or performance improvement is claimed by this publication.
Review notes
The prerequisite PRs remain separate review units. This draft records their composition fixes and tests; do not apply both its aggregate patch and the original stacks twice. Rebase/narrow the diff after prerequisites land. An open-PR search found the existing measurement/runtime proposals, not a separate implementation of these composition fixes.
AI assistance: OpenAI Codex. Keep Draft until human review and the outstanding production checks are complete.