GitLab

* Refactor to allow more flexible treatment of multiple optimizers/models/losses

* Adding _process_optimizers.py

* Created L0 tests (now passing).

* fix: minor print typo (#234)

* make L1 results easier to read

* L0 multiple model/optimizer/loss test fleshed out

* Adding test that master params remain synced across distributed processes

* Docstring updates

* Docstring updates

Fix dispatch where we have a parameter group with multiple
combinations of types
Optionally apply weight decay after momentum

Incorrect types used in a few places

Only support the 4 specific cases we care about
Remove more general set of switch statements

Fuse in fp16 gradient -> fp32 convert
Additional option fp16 weight copy written out

Initial implementation, all fp32
Tested against torch.optim.sgd

Allow multi-tensor unscale to handle FP16 output, so it can also be used for copy-scatter. Rename some options.

Added kernel to support sync BN for channel last tensor