[docs] Concept guides

[stevhliu]

This PR focuses on updating the concept guides with additional PEFT methods. Among other things, it also:

• reorganizes the concept guides into broader categories like "Reparametrization" (to include LoHa, LoKr, OFT, and AdaLoRA) and "Soft prompts"
• adds a separate developer guide for using LoRA because some of the content in the LoRA concept guide is instructional; expands a bit more on the LoRA utility methods

todo:

• add multitask prompt tuning
• add Llama-Adapter

[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]

@stevhliu Let us know when this is ready for review.

[BenjaminBossan]

Thanks so much for this work, this must have been a lot of effort and it fills a big gap we currently have in the docs.

I made a couple of suggestions, please check if they make sense to you.

[BenjaminBossan]

I wonder if we could cause confusion here. We use the name "adapter" pretty liberally throughout PEFT for all kinds of methods. Users may, e.g., pass an adapter_name to a LoRA model. This section makes it seem that LLaMA adapter is the the only "adapter" method in PEFT.

Honestly, I'm not sure how to correctly categorize this method, there is no clear taxonomy of adapter/tuner methods :-/ But I'd rather combine this with lora.md and call all of them adapter, as "reparametrization" is not a name I have seen used colloquially for these methods.

[stevhliu]

I was struggling a bit with the taxonomy of it all as well, but I think your approach is better for avoiding confusion! 👍

[younesbelkada]

Impressive work @stevhliu ! Thanks very much for all your hardwork on this!
I left one tiny suggestion, what do you think?

[pacman100]

Wow! Thank you @stevhliu for updating the docs with overview of all the methods, usage of the features and pointers to the related resources! 🤩🚀✨

[BenjaminBossan]

Looks great, thanks for the updates.