Created by: stephenroller
Patch Description Recent changes to dataloaders have resulted in us creating checkpoint files that are more unequal in size, as shard0 contains the data state. This breaks this older heuristic we used to determine if a checkpoint was Zero2 or Zero3.
This patch changes the heuristic to just read from the config of the file itself.
This patch also gives a better error message for a common mistake.
Testing steps Internal evaluations.