Coalesce packed-expert optimizer checkpoint copies - #979
Draft
bradhilton wants to merge 2 commits into
Draft
bradhilton wants to merge 2 commits into
bradhilton wants to merge 2 commits into
Conversation
bradhilton
deployed
to
trainer-rank-gpu-validation
September 26, 2026 03:13 — with 
GitHub Actions
Active
GitHub Actions
Active
bradhilton
had a problem deploying
to
trainer-rank-gpu-validation
September 26, 2026 08:15 — with
GitHub Actions
Failure
GitHub Actions
Failure
This branch had an error being deployed
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Packed-expert checkpoint saves currently perform three blocking optimizer CPU transfers per exported expert. Copy each complete consecutive packed parameter once, then slice its owned CPU storage. Sparse, reordered, interleaved and mismatched-shape exports retain the existing slice path.
The preparation fence, synchronous snapshot writes, checkpoint format and finalize/abort order stay unchanged. An eight-expert fixture reduces optimizer transfer calls from 48 to 6 without changing transferred bytes. This uses additional temporary packed CPU storage; GPU latency and peak memory have not been measured.
Validation: 33 CPU unit tests pass for dtype/layout, missing moments, copy counts, snapshot ownership, cancellation cleanup and concurrent preparation. The independently reviewed isolated component also passed 27 differential controls, including exact finalized checkpoint bytes and manifest digest. Ruff lint/format and targeted fixture type checking pass. Canonical quality CI passes at the current head, including lint, format, typecheck, lock, lightweight and unit tests. Its GPU job could not acquire two H200s, so it produced no new GPU test result.
Draft for review only; no frozen-run changes or deployment.