Continue MPS batches after a failed seed exhausts retries - #58
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Purpose
Prevent one transient Apple-MPS/Metal failure from terminating a multi-seed or multi-optimizer batch after the existing fresh-process retry is exhausted.
Behavior
The default MPS policy is now:
The failure is not hidden:
run_failed.json;_batch_status.json;PARTIAL SUCCESSsummary;run_complete.jsonis written for the failed replicate, so completed-run analysis continues to exclude it.--fail-fastremains available for debugging and restores immediate termination after retries are exhausted.If a worker fails before any finite checkpoint is written and another retry remains, the partial run directory is archived and the fresh worker restarts from initialization instead of attempting an impossible resume.
Scientific invariants
This does not change model equations, Muon or MuonClip updates, LR schedules, data sampling, evaluation probes, WeightWatcher settings, checkpoint selection, seeds, or result layout. It only changes batch-level error handling.
Tests
The expanded MPS recovery tests verify:
_batch_status.jsonaccurately reports partial success;run_failed.jsonrecords the incomplete replicate;--fail-faststill stops immediately when requested.