ENH Work on supporting scikit-learn intelex

[BenjaminBossan]

Partially solves #251

This is one part of the work required to solve the mentioned issue. The other part will have to be added on the API inference side, once this PR is finalized.

Description

Added the option use_intelex to hub_utils.init. It adds a new entry to config.json which can later be used by the inference API to decide whether to run it with intelex or not.

On top of that, if use_intelex=True, scikit-learn-intelex will be added as a dependency to the requirements if not already there. Moreover, if metadata for a model card is loaded, a "scikit-learn-intelex" tag will be added.

[BenjaminBossan]

@skops-dev/maintainers Ready for review.

[adrinjalali]

Thanks @BenjaminBossan

[BenjaminBossan]

@adrinjalali I addressed your comments, pls review again.

[adrinjalali]

I just saw this:

https://github.com/intel/scikit-learn-intelex/blob/d71f63c12cba7ee99191e3013182b0650ac9ec71/sklearnex/dispatcher.py#L104-L107

def patch_sklearn(name=None, verbose=True, global_patch=False):
    if not sklearn_check_version('0.22'):
        raise NotImplementedError("Intel(R) Extension for Scikit-learn* patches apply "
                                  "for scikit-learn >= 0.22 only ...")

Which patching wouldn't be an option for us. That's a VERY OLD sklearn version. I checked their codebase, and they have their own estimators which can be imported and used using explicit imports from their library.

If the user explicitly imports from the scikit-learn-intelex library, then everything works as expected. They can have it as an extra dependency, and the backend would install it and load the model successfully.

Which means this task would be reduced to a documentation and an example script on our side, to showcase how one can use scikit-learn-intelex on HF hub. WDYT @BenjaminBossan @echarlaix

[BenjaminBossan]

Which patching wouldn't be an option for us. That's a VERY OLD sklearn version.

I think this is the minimum sklearn version, so we should be good. When I tested locally, patching worked (or, at least, didn't raise an error).

[adrinjalali]

Oh sorry, it's checking if the sklearn version is >0.22, that's a very bad way of writing code lol, but sure, it works.

[BenjaminBossan]

Oh sorry, it's checking if the sklearn version is >0.22, that's a very bad way of writing code lol, but sure, it works.

Yes, naming is hard :)