Support for Custom Adapters

[dgme-syz]

Feature request

In simple terms, I would like support that allows users to customize their own adapter. I noticed that users only need to add a folder under this path src/peft/tuners and place some adapter-related files, usually config.py, layer.py, and model.py.

However, during implementation, I found that I also need to modify src/peft/utils/save_and_load.py/get_peft_model_state_dict to ensure that the custom adapter can be saved correctly. This is because the function is currently only adapted for existing adapters, so I have to modify the source code to ensure that the custom adapter can be used successfully.

PEFT is the most convenient and efficient fine-tuning library, and it would be even better if this feature were supported. Perhaps you’ve already implemented this functionality, but I haven’t found it yet. If so, please point it out. Thank you very much.

Motivation

I hope to use custom adapters to fine-tune large language models.

Your contribution

Currently, I have no clear ideas.

[BenjaminBossan]

You are correct, when adding a new method, besides creating a new directory inside of src/peft/tuners, there is also the necessity to make a few adjustments to other PEFT files (usually very small adjustments, but it depends on the type of adapter).

Regarding the change in get_peft_model_state_dict, yes, we can refactor this to work out of the box without further changes for new adapters, I'll take a look. But note that by adding new code into src/peft/tuners, you're already making changes to the PEFT code base, so editing get_peft_model_state_dict is not really that different.

Perhaps you’ve already implemented this functionality

No, but as mentioned, I'll look into this, I had planned to facilitate this for a long time. The only thing we have right now is a way to add new custom LoRA layers, but not yet to add completely new adapter types.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

[BenjaminBossan]

not stale, see #2282

[BenjaminBossan]

@dgme-syz The PR is merged, adding a new PEFT method is now much simplified. To quote myself:

Ignoring tests, docs, and examples, we have the additions to src/peft/tuners/my_peft_method, but also need to:

1. Add an entry to PEFT_TYPE_TO_CONFIG_MAPPING in mapping.py.
2. Add an entry to PEFT_TYPE_TO_TUNER_MAPPING in mapping.py.
3. Add an entry to PEFT_TYPE_TO_MODEL_MAPPING in peft_model.py
4. Add an entry to PEFT_TYPE_TO_PREFIX_MAPPING in utils/constants.py
5. Add some code to get_peft_model_state_dict in utils.save_and_load.py

With the changes in this PR, all these steps can be omitted.

On top of that, we also have the re-imports to peft/__init__.py and peft/tuners/__init__.py but those are still required (I'm hesitant to mess with the import system). Furthermore, it's still required to add an entry to PeftType in utils.peft_types.py. Since this is an enum, it can't be easily generated automatically. Therefore, adding a new PEFT method is still not 100% self-contained.

To register a new PEFT method, you should now add this to peft/tuners/my_peft_method/__init__.py:

from peft.utils import register_peft_method

register_peft_method(name="my_peft_method", config_cls=MyConfig, model_cls=MyModel)