Refactored auto-microbatching hook handles for FSDP

[rithwik-db]

Refactored auto-microbatching hook handles for FSDP1 with additional documentation.

This PR was originally designed to support FSDP2 auto microbatching, but since there are additional issues with FSDP2 state there, we moved that to a draft PR: #3866

[bowenyang008]

can we run an e2e test to verify it works?

[rithwik-db]

can we run an e2e test to verify it works?

It seems hard to curate an e2e test to catch this failure that would be more informative than our current unit tests. We have these two unit tests:

1. FSDP1
2. FSDP2

I guess in theory, we could create a larger example where a certain MPT module raises a CUDA OOM error at a certain epoch given a specific batch size or we could use a MPT module with a massive hidden layer in the FFN that will run into OOM for one batch size but not 1/2 of it...

[bowenyang008]

can we run an e2e test to verify it works?

It seems hard to curate an e2e test to catch this failure that would be more informative than our current unit tests. We have these two unit tests:

1. FSDP1
2. FSDP2

I guess in theory, we could create a larger example where a certain MPT module raises a CUDA OOM error at a certain epoch given a specific batch size or we could use a MPT module with a massive hidden layer in the FFN that will run into OOM for one batch size but not 1/2 of it...

I meant just a general e2e test, does not have to trigger OOM

[rithwik-db]

Tested here: mpt-7b-fsdp2-p39FPR and compared it to base (mpt-7b-fsdp2-AKLNwv) and the numbers look good based on the tolerations mentioned in the regression testing PR @bowenyang008 (note that it defaults to 8 microbatch size when auto is set)