mosaicml / composer

Supercharge Your Model Training
http://docs.mosaicml.com
Apache License 2.0
5.12k stars 415 forks source link

Remove save overwrite #3431

Closed mvpatel2000 closed 3 months ago

mvpatel2000 commented 3 months ago

What does this PR do?

Previously, we required save_overwrite=False as it was not clear if a user intended to overwrite or resume from a checkpoint.

However, it turns out this situation is well defined. Autoresume = resume from symlink file. It is a transformation on load_path, and nothing else. So, you can actually set autoresume on and save_overwrite to True, which would be useful for partial checkpoint uploads where the symlink file is not added.

mvpatel2000 commented 3 months ago

Need approval from either Ning or Evan. Did we not test that save_overwrite=True and autoresume=True threw an error before?

We did, I removed test

eracah commented 3 months ago

oh too late