Make Jsonformer derive from PreTrainedModel

[fkarg]

As far as I understand it, this is currently not the case.

Having Jsonformer derive from PreTrainedModel would enable immediate use with e.g. pipeline and other ecosystem building blocks that require a PreTrainedModel.

It might even be possible to automatically derive directly from the (more specialized) base class loaded using from_pretrained (and automatically loading the tokenizer from the same path except specified otherwise). That way, almost no functions would need to be changed. Other ideas:

• implement forward manually. This is probably tedious
• automatically load all functions from the other model and set them, e.g. along the lines of

from inspect import getmembers, isfunction

    # somewhere in Jsonformer.__init__, probably
    for name, func in dict(getmembers(self.model, isfunction)):
        setattr(self, name, func)

Edit: Thinking about it some more (and understanding the PreTrainedModel interface better), it's probably not that easy.