Hanlin Tang
Hanlin Tang
Layer freezing currently doesn't support resuming from checkpoints because of circular requirements: 1. similar to model surgery, layer freezing modifies the optimizer param groups, and therefore needs to be applied...
Users running locally may want a TB or ML-flow logger backend.
Based on feedback from @tcapelle, our `wandb` logging experience can be improved with better naming, and more default things logged, see comment here: > We could improve the naming of...
Tests are good code snippets for how to use our trainer. However, most of our tests of trainer functions (e.g. checkpointing, DDP) all rely on the hparams system, which makes...
Based on this discussion (https://github.com/mosaicml/composer/pull/393#discussion_r803090943), the current `@device(, precision=True` decorator is a little too complex and multipurpose. A few proprosals have been made for a simpler design, but those need...
Current testing for streaming dataloader, e.g. C4, should be used with mock sources to avoid streaming from buckets directly in CI.
This PR removes YAHP from our codebase. Making this PR early to suss out any issues, and form a base for testing. Do not merge this PR until after the...
This is the first in a series of PRs that brings this library into compliance with `pyright`. No functional changes to the code should occur with these fixes. Before: ```...