[fix] revision of the adapter model can now be specified.

[pesuchin]

Resolved: #3061

Solution

I modified the _load_peft_model method so that when the config is PeftConfig, it handles loading both the base model and the adapter model. Additionally, when loading the base model, I temporarily save and remove the revision from model_args before loading the base model.

This ensures that the base model is loaded using the default revision instead of the adapter model's revision, resolving the error.

Verification

The following three codes were tested and worked correctly.

• verification1:

from sentence_transformers.models.Transformer import Transformer

args = {
    "token": <your token>,
    "trust_remote_code": False,
    "revision": <your repository revision>,
    "local_files_only": False
}

transformer = Transformer(
    model_name_or_path=<your model_path>,
    cache_dir=None,
    backend="torch",
    max_seq_length=512,
    do_lower_case=True,
    model_args=args,
    tokenizer_args=args,
    config_args=args
)

• verification2:

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("sentence-transformers-testing/stsb-bert-tiny-safetensors")
model.encode("Hello, World!")

• verification3:

from sentence_transformers import SentenceTransformer

kwargs = {
    "revision": <your revision>,
}
model = SentenceTransformer(<your model repository path>,
                            use_auth_token=<your token>,
                            model_kwargs=kwargs,
                            config_kwargs=kwargs,
                            tokenizer_kwargs=kwargs)
model.encode("Hello, World!")

[tomaarsen]

Hello!

I see a minor issue here that I'd like to see resolved:

• The T5 and MT5 models can also have PEFT Adapters, but no longer with this PR.

I had a crack at this, and I think the easiest solution is actually to have the _load_config return whether the model is a PEFT Model or not. That simplifies everything else a lot. Then we can load the base model like before, but beforehand we update the model_args and then afterwards we load the PEFT adapter on top.

I also added a simple test with the https://huggingface.co/sentence-transformers-testing/stsb-bert-tiny-lora model.

I'm curious to hear what you think!

• Tom Aarsen

[tomaarsen]

Apologies for expanding the PR a bit more via 7ee9fa8. In https://github.com/UKPLab/sentence-transformers/pull/3079/files you can select only the first 2 commits of this PR by shift clicking the first 2

That's the easiest way to understand my changes in ec366b5

• Tom Aarsen

[pesuchin]

@tomaarsen
Hello!
Thank you for your suggested revisions.
I think it’s a great idea.
The implementation has become much simpler.
Thank you for adding a simple test as well.