[enh] Add Support for multiple adapters on Transformers-based models

[carlesonielfa]

Hi!

Pull Request overview

Adds support for loading, activating and deactivating adapters for a SentenceTransformer model when its underlying model is Transformers-based.

Details

Adds a wrapper mixin so that the methods to manipulate adapters are exposed on the SentenceTransformer model. It includes checks to ensure that the underlying model supports it.

I wanted to implement it in a manner that won't break or require maintenance if the transformers library changes the signature for these methods, this however compromises typing and hints, so if that is more of a concern than maintenance, the implementation could be changed to a more hard-coded one.

As it is only a wrapper, the functionality follows that documented on the hf docs for transformers .

Usage example

Pre-trained adapter

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("my_org/mymodel")
model.load_adapter("my_org/my_pretrained_adapter", "my_adapter")

Adapter from scratch for training purposes

from sentence_transformers import SentenceTransformer
from peft import LoraConfig, TaskType

model = SentenceTransformer("my_org/mymodel")

peft_config = LoraConfig(
    target_modules=["query", "key", "value"],
    task_type=TaskType.FEATURE_EXTRACTION,
    inference_mode=False,
    r=8,
    lora_alpha=32,
    lora_dropout=0.1
)
model.add_adapter(peft_config)

Additional considerations

• Consider incorporating adapter switching to the SentenceTransformer.encode function
• Test with a pretrained model is commented out on this PR waiting for feedback from a maintainer on if an adapter model should be uploaded to the sentence-transformers hf hub
• Might be worth including documentation about the feature