fix from_pretrained in offline mode when model is preloaded in cache

[oOraph]

What does this PR do?

Fixes # (issue)
fix model load in offline mode

Who can review?

@ArthurZucker
@Wauplin

Reproduce:

1. download a private model from hub
2. script:

import os

os.environ['TRANSFORMERS_OFFLINE'] = "true"

from transformers.models.auto.modeling_auto import AutoModelForImageClassification

model_id = "Narsil/private"
revision = "main"
kwargs = {'use_auth_token': None}


model = AutoModelForImageClassification.from_pretrained(model_id, revision=revision, **kwargs)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

Thanks! Can you add this example by using HF hub to downlaod snapshot of a .bin file in a temp repo? and test both the flag to go offline and the local_files_only in from pretrained? 🤗

[Wauplin]

Thanks @oOraph!

[oOraph]

Thanks! Can you add this example by using HF hub to downlaod snapshot of a .bin file in a temp repo? and test both the flag to go offline and the local_files_only in from pretrained? 🤗

Done @ArthurZucker , added some unit test -> fails before fix, succeeds after
test is ugly but I could not find any cleaner way to do it...

[LysandreJik]

Thanks!

[ArthurZucker]

https://huggingface.co/hf-internal-testing/tiny-random-SiglipForImageClassification/tree/main does not have a bin file, so we are not testing the case we solved before. Otherwise LGTM