Add retry hf hub decorator

[muellerzr]

What does this PR do?

This PR adds a @hub_retry decorator, which should be used on tests that give us any form of requests-like error when a download from the hub fails for any reason. It'll then wait a beat then try to rerun it again. I currently have it just located in the
ModelTesterMixin class, but as time goes by and see where other places have it pop up, we should use it in those mixins/classes as well.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@ydshieh

[BenjaminBossan]

This is a useful feature. I wonder if there is a possibility to use this for monkey patching from_pretrained. That way, it would be easier to use the retry mechanism in other packages. As an example, in PEFT we don't have a base class for all tests, so the trick with __init_subclass__ would not work.

[ydshieh]

@BenjaminBossan I was also thinking if we can instead patching from_pretrained in conftest.py rather than this change in __init__, but I can's judge it's better than the current version.

[ydshieh]

Also, it looks this could be applied to any function (test, from_pretrained) directly, so

monkey patching from_pretrained

is just hub_retry(XXX_class.from_pretrained) maybe

[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.

[BenjaminBossan]

is just hub_retry(XXX_class.from_pretrained) maybe

Right, this could work. Do you know if there is a single method that can be patched that would work for all transformers auto models?

[ydshieh]

Not a single method though.

from_pretrained for _BaseAutoModelClass will work for all auto model, but there are also auto tokenizer/processors etc.
And There are also PreTrainedModel if auto is not used (then similarly for tokenizer/processors etc)

[BenjaminBossan]

Wondering if it's better to use exponential backoff instead of a fixed waiting time?

[muellerzr]

We can see if fixed gets the job done before overengineering? :)

[muellerzr]

Also @Wauplin may just put this in huggingface_hub directly so this might (hopefully) just be temporary

[Wauplin]

maybe at some point yes but not short-term so please go ahead with this PR :) (cc @hanouticelina for viz')

[muellerzr]

For now going with what's in here (to just get it merged)

[BenjaminBossan]

Thanks for picking this back up. LGTM.

[ydshieh]

Thank you for finalize this! Hope our life is easier with this 🙏

The test is written in a class that require torch, but that test is super simple not need torch. We can move it later, no big deal.

Let's merge and see how it goes 🚀