Created by: stephenroller
Patch Description #189 added support for subshards, which is important for creating smaller shards (virtually) and allowing faster rewinding of dataloaders. However, subshard values that are reasonable for training (e.g. 50) are not reasonable for validation, as the validation set is much smaller.
This patch makes it so subshards are only enabled in training, and ignored elsewhere.
Testing steps
Ran a training run with --data-subshard-count 50