[PEFT] Pass token when calling find_adapter_config

[younesbelkada]

What does this PR do?

This PR fixes an issue that was reported on Spaces. I was not able to reproduce the issue locally though.
When loading a model with token=True (i.e. on a gated or private repository), find_adapter_config will try to look for an adapter file inside a private repository without the token, leading to an authentication error.

The fix is to pass the token to adapter_kwargs and remove the duplication here: https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_utils.py#L2529

cc @LysandreJik

[HuggingFaceDocBuilderDev]

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

[abhishekkrthakur]

This seems to fix the issue for me. Shall we merge it? Or figure out the actual cause?

[LysandreJik]

LGTM

[denys-fridman]

@younesbelkada adapter_kwargs can be None, in which case the code breaks. Here it can be set to None, then it passed via this line.

For reference, I was running

python main.py \
  --model path_to_mistral_7B \
  --tasks mbpp \
  --temperature 0.1 \
  --n_samples 15 \
  --batch_size 4 \
  --precision fp16 \
  --allow_code_execution \
  --save_generations \
  --max_length_generation 512 \
  --save_generations_path predictions.json \
  --metric_output_path metrics.json

[younesbelkada]

Nice catch! Will open a PR soon to fix that