Skip to content
GitLab

Automatically evaluate checkpoints after copying to NFS

Merged Administrator requested to merge auto_eval into main

Created by: Xirider

This PR adds the ability to automatically evaluate checkpoints after copying to NFS.

  • After the asynchronous copy of a checkpoint to NFS, we check periodically whether all checkpoint parts have finished copying.
  • Once all parts are done, we run the eval script defined in nfs_eval_script_path for this checkpoint

Add these lines to the internal sweep script to define the eval script path: https://www.internalfb.com/phabricator/paste/view/P576097431

Testing: I tested this with a dummy eval script, as the Punit's real eval script should only be used for large runs. I would recommend to check first whether the eval script alone runs correctly on our frozen branch.