Repository navigation
synmon: resume swe-gym from checkpoints after trainer preemption - #303
Draft
Micah Yong (micahtyong) wants to merge 1 commit into
Draft
Micah Yong (micahtyong) wants to merge 1 commit into
Micah Yong (micahtyong) wants to merge 1 commit into
Conversation
…tor nonpreemptible Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch has not been 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
The 2026-11-16 nightly swe-gym run failed after two preemptions:
forward_backwardreturned HTTP 410model_lost.Losing models on a GPU preemption is expected. The swe-gym example already saves state after every step and can continue with
--resume. This PR makes the nightly harness use that path:monitoris now nonpreemptible, so a client run no longer restarts from scratch.model_lost, the harness reruns the stock example once with--resume, using the remaining time budget. Each client continues from its last checkpoint. Other failures still fail the night.E2E
model_losterror, exits 1, and then finishes its remaining steps on--resume. The harness resumed once, counted 5 steps per client, and reported "resumed from checkpoints, not compared".exit 1: helpers.environments.InfrastructureError: Nonfinite training metrics.🤖 Generated with Claude Code