add example and update deepspeed/FSDP docs

[pacman100]

What does this PR do?

1. Add SFT example and update DeepSpeed/FSDP docs.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[BenjaminBossan]

This is a great guide, thanks a lot Sourab. I already took a look at it, not sure if it's ready for review yet :)

Apart from the comments I left, do you know some common mistakes that are easy to make when using PEFT + DS? Or some PEFT features that won't work when combined with DS (like merging weights)? Maybe those could be added to the troubleshooting section and a link added in this guide.

[BenjaminBossan]

Let me know once this is ready for another review round.

[BenjaminBossan]

Thanks a lot for extending the PEFT docs on DeepSpeed and FSDP. Your additions are super good 🔥🔥🔥

Also very nice that you added the example scripts to along with the docs. Some comments on those:

• Is examples/sft/__init__.py needed? I think it can be deleted.
• Pretty cool that you also included the multi-GPU and unsloth script. Right now, they are not referenced in the docs. Maybe a short README.md could be added to the folder to explain the individual scripts, just very briefly.

I also added some small comments to the docs, but those are just nits, not blockers. Do we also want to ask Steven for a review?

[pacman100]

Pretty cool that you also included the multi-GPU and unsloth script. Right now, they are not referenced in the docs. Maybe a short README.md could be added to the folder to explain the individual scripts, just very briefly.

Done.