Skip to content

Resume filter validation from its own CSV after a crash - #451

Open
luciaquirke wants to merge 1 commit into
mainfrom
filter-crash-resume
Open

Resume filter validation from its own CSV after a crash#451
luciaquirke wants to merge 1 commit into
mainfrom
filter-crash-resume

Conversation

@luciaquirke

Copy link
Copy Markdown
Collaborator

A filter validation run is one from-scratch retrain per query — hours of work at larger corpora — and the per-query CSV is already written incrementally, but an interrupted run had no way to continue: with resume the query loop restarted from zero, and without it the existing run_path is refused entirely.

This reads the method CSV back on startup, re-emits the measured rows, and skips their retrains. The CSV is read on every rank (not just rank 0) so the distributed retrain loop stays in step. A crash now costs one query's retrain instead of the whole run; measured under repeated shared-volume outages, where each interruption previously discarded up to 2.5 h.

🤖 Generated with Claude Code

https://claude.ai/code/session_0167XKnE2yoCE162n3jHvxC1

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant