PoC of getting model_class from the Hub instead of having to pass a --feature

[julien-c]

related to internal discussion: https://huggingface.slack.com/archives/C03AHFPCWHM/p1662407936101009

[hollance]

If we're going to go this route, I'd rather get rid of the feature argument altogether and not have two ways to achieve the same thing. But that's something we should really do together with the ONNX exporter.

[julien-c]

+1 on removing the feature argument, and IMO it could be replaced by a model_class flag or similar, in cases where one wants to override the architecture

Re. ONNX, yes possibly we could do it in both at the same time?

[hollance]

One thing I want to do first is implement the with-past logic. Apparently this was one of the reasons for adding feature and it will affect some of the decisions we might make around this issue.

[julien-c]

sounds good!

happy to share more info or doc about what's available on the Hub side in terms of transformers models metadata. cc @sgugger who's been working on this and knows how it works