allenai / OLMo

Modeling, training, eval, and inference code for OLMo
https://allenai.org/olmo
Apache License 2.0
4.2k stars 392 forks source link

Unshard without passing checkpointer type #603

Closed 2015aroras closed 1 month ago

2015aroras commented 1 month ago

The unsharding script currently requires us to pass in the checkpoint type, but this is not necessary since it can be read from config. Passing in the type has been a nuisance in the past, so this PR changes the script to default to the checkpoint type in the config. It changes the unsharder to use build_sharded_checkpointer, so that we don't need to duplicate that logic in the unsharding script.

Currently untested.

2015aroras commented 1 month ago

This worked fine for a checkpoint created with the local checkpointer.