Hanlin Tang

Looking at `torch.fx`, they do support graph re-writes and even subgraph pattern matching (https://pytorch.org/docs/stable/fx.html#torch.fx.replace_pattern). Looks like they return a `torch.nn.Module`, but we'll need to check its compatibility with other surgery...

Sounds good, looking forward to collaborate! We care most about delivering efficiency methods to the community in whichever form they would like to consume.

I think using `composer -n 1` is basically saying using `torch.distributed`, so maybe?

Ah, I missed the `README` getting deleted. I think we definitely want `README` for the `algorithms` and `model` folders, as they show up when users are navigating github. It's OK...

@A-Jacobson some conflicts to resolve, but otherwise ready to merge

Thanks @JJGO , agreed we'll add the API docs to the method cards, that's' probably the easiest entrypoint. Unfortunately the nested `composer.algorithms.blurpool.blurpool` is difficult to solve unless we either (1)...

@mvpatel2000 I think you can replicate the issue if you try: `pip install --upgrade datasets` (to >=version 2)

yeah agreed this is a bit annoying when trying to look at different method cards, as each time i click on one, the scroll bar goes to the top.

Closing this -- please re-open if you encounter the same issue!

This does seem extraneous as `model_inputs['labels']` is not used anywhere.. @moinnadeem can you confirm?