Repository navigation
fix(ppo): preserve best metric across regular checkpoint resume - #1333
Merged
hijkzzz merged 1 commit intoSep 12, 2026
Merged
Conversation
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Regular PPO checkpoints omit the historical best evaluation metric. Resuming from one resets the best threshold to
-inf, so a worse subsequent evaluation can be saved as the new best and delete the previous best checkpoint.Persist best metric key/value in regular checkpoint client state and retain the latest evaluation value separately for checkpoint rotation. The existing restore path accepts older checkpoints using its previous fallback. Both synchronous and asynchronous PPO use these base methods.
Validation: full pytest suite (70 passed), compileall and pre-commit. Regression tests cover save/load, worse and better evaluations, zero/negative metrics and older metadata. A CPU file-system probe ran the actual trainer decisions and strategy checkpoint rotation: the old best survives a worse evaluation after resume and is replaced by a real improvement. Ray transport and DeepSpeed engine serialization were stand-ins; no GPU training was run.