Enable PeftConfig & PeftModel to load from revision

[lewtun]

This PR enables the from_pretrained of PeftConfig and PeftModel to load configs/models with a specific revision arg. This is useful when storing many adapter weights as branches in a model repo.

Sample usage

from peft import PeftConfig, PeftModel
from transformers import AutoModelForCausalLM

model_id = "lewtun/tiny-random-OPTForCausalLM-delta"
revision = "v1"
config = PeftConfig.from_pretrained(model_id, revision=revision)

model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
)
model = PeftModel.from_pretrained(model, model_id, revision=revision)
assert isinstance(model, PeftModel)

[lewtun]

If you're happy with this approach, I'll move the model to hf-internal-testing

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[lewtun]

Note that this repo only has weights in the v1 branch to properly test the revision arg works as expected

[pacman100]

Thank you @lewtun for improving the load and save utils! 🤗

[winglian]

@pacman100 This causes a regression when device_map is used for PeftModel.from_pretrained. Probably not a great idea to pass all **kwargs blindly from PeftModel to PeftConfig.