Improve FSPD

[schmidt-ai]

This PR attempts to decouple model training from FSDP, allowing more flexible distributed training e.g. with pytorch-lightning.

IIUC we should be wrapping Block, not BlockChunk (which is the same as torch.nn.Sequential), but if that's incorrect we can revert that change.

[facebook-github-bot]

Hi @schmidt-ai!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!

[qasfb]

So the point of wrapping around BlockChunk instances is for avoiding the overhead of small collective ops during FSDP training (if FSDP instances are the blocks), or long idle waits (if the whole model is an FSDP instance; it needs to gather it entirely before starting computation), for me it is a parameter that affects iteration time.