Changes suggested to fastai integration

[omarespejel]

A cluster of changes suggested to the fastai integration after the merge of #678

• Instead of moving the fastai.Learner from the directory, save in current directory (see comment in Add fastai upstream and downstream capacities for fastai>=2.4 and fastcore>=1.3.27 versions #678). This is an issue with how fastai exports the Learners.
• Consider using packaging and packaging.markers when reviewing if the local fastai and fastcore versions are supported (see comment in Add fastai upstream and downstream capacities for fastai>=2.4 and fastcore>=1.3.27 versions #678).
• Right now we require that a pyproject.toml file exists inside the repo of a fastai model. We might not want to be super strict. Maybe if there is no pyproject.toml or no versions specified there, we can just show a warning instead of breaking entirely.

[adrinjalali]

The third issue is related to the backend and not the integration, is that correct?

The first and second issues I think we need to fix in the existing PR since changing them is a change in the behavior for the users.

[osanseviero]

The third point means that instead of raising errors when loading a model if there is no toml specified, we could instead just load the model but show a clear warning to users.

Pretty much making these less strict

• https://github.com/huggingface/huggingface_hub/pull/678/files#diff-21cb622665e96594b1761f0db9fa7465e605d4e2a12d8e1c080c5de597738423R101-R108
• https://github.com/huggingface/huggingface_hub/pull/678/files#diff-21cb622665e96594b1761f0db9fa7465e605d4e2a12d8e1c080c5de597738423R112-R130

I suggested Omar we can do this particular point in a follow-up PR since I think it's better to get this out and get some feedback from users that will be using this from head (not from release) so we can also do other changes based on the feedback.

I agree the first point in particular should be fixed in the existing PR. The second seems like a minor internal implementation thing that should not change user behavior, so no strong opinions.

[omarespejel]

Staling for the moment. Lacking point two: Consider using packaging and packaging.markers when reviewing if the local fastai and fastcore versions are supported.

[adrinjalali]

Would you mind letting us know why it's closed as won't fix?

[omarespejel]

@adrinjalali, I was staling it since it's not on the short-term roadmap to fix it. However, I reopened it now for future reference.

[osanseviero]

That's not a reason to close an issue though. An issue documents a feature/thing we want to add/change in this repo, both in short and long term. Even if it won't be worked in right now, it's still good to have. If the issue is closed it will likely never be revisited 😄