Refactor layer initialization to use PEFT config directly

[BenjaminBossan]

Description

There is a lot of code duplication in the way that adapter layers are initialized. This means that if, say, a new LoRA config argument is added, we need updates in the following places:

1. LoraModel._create_and_replace: kwargs dict
2. LoraModel._create_and_replace: update_layer arguments
3. LoraModel._create_and_replace: update_layer signature
4. Linear.__init__: signature (same for each LoRA layer type)
5. Linear.__init__: update_layer arguments
6. resolve_lora_variant: signature if the argument is tied to a LoRA variant

It is far too easy to forget something and it's in fact a common mistake we see, especially by outside contributors.

This PR aims at removing this duplication as much as possible. The crux is that almost all of these arguments derive from the LoraConfig. Therefore, the duplication can be reduced by passing the LoraConfig directly instead of each of its individual arguments.

As is, this change is only implemented by LoRA, which is by far the most affected by this duplication. However, an argument could be made that for consistency, the other PEFT methods should be refactored accordingly. Since AdaLoRA inherits from LoRA, it was also adapted.

Possible counter points

With this change, the used arguments are not listed explicitly anymore. But since the LoraConfig is a typed dataclass, this replacement does not stop us from knowing which arguments are available (unlike, say, using untyped kwargs). So for editing support and readability, this replacement should not hurt.

Another possible issue with this proposal is that it could easily break some functionality, as it can be easy to overlook something. Not everything is covered by tests (some of the more rarely used quantized layer types, TP layer), so it's possible to miss this.

[BenjaminBossan]

Note: I added variables like these to reduce diff noise. LMK if you'd rather remove those and just use config.lora_dropout etc. below.

[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]

Note: r and alpha can differ from the one in the config (e.g. because of rank_pattern), thus they are not taken from the config.

[githubnemo]

I think this looks fine but I noticed that lora/gptq.py seems unmodified still, is this on purpose?

[BenjaminBossan]

@githubnemo Thanks for pointing this out, GPTQ is now also updated. I ran the GPTQ tests locally (as they are GPU tests) and they passed.

[githubnemo]

LGTM, thanks

[BenjaminBossan]

@hiyouga Just a heads up, this PR simplifies the signature of multiple methods like update_layer. The idea is to directly pass the PEFT config instead of unpacking the arguments, which saves us a lot of code and makes PEFT easier to extend. For now, this is only for LoRA, but it will be extended to other PEFT methods later. In cause you're overriding these methods, please take a look at the new signature. From skimming LlamaFactory, it looks like you should be safe, but please check your tests using the PEFT main branch to be safe. If you encounter issues, please let us know.

[hiyouga]

@BenjaminBossan Thanks for the heads-up! All unit tests in LlamaFactory pass with the PEFT main branch, and we haven't encountered any issues.