Add --resume: one command converges any fan-out run
#1
by davanstrien HF Staff - opened
Recovery no longer requires knowing which rank failed:
--resume --run-id <id>: waits for any still-in-flight workers, diffs the
manifest against worker done-states in the bucket, re-runs exactly the
missing ranks, then consolidates. Idempotent - safe to run repeatedly.- Attached runs now auto-retry failed workers once before giving up, and the
failure message points at--resumeinstead of per-rank instructions. - README: recovery section + fleet dashboard link.
Tested on Jobs: 2-shard run launched with --no-wait, one worker canceled
mid-run, single --resume converged it to the exact 4,000-row output
(davanstrien/resume-smoke-test).
Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Superseded by the source-of-truth GitHub PR: https://github.com/davanstrien/uv-scripts-for-ai/pull/87 (this repo is mirrored from uv-scripts-for-ai via the sync workflow — changes land there first). Closing.
davanstrien changed pull request status to closed