v0.1.3
Optimizers
- Adds Sophia & StableAdam optimizers
- Adds native fastai support for bitsandbytes 8-bit optimizers
- Reduce ForEach L2 weight decay and RAdam, LAMB, & Ranger optimizer memory usage
- Increase LAMB step speed
Other Features
- Add asynchronous fastai batch transforms to the FFCV Loader
- Add DynamoExplain callback for diagnosing
torch.compile
results
Full Changelog: v0.1.2...v0.1.3