Other computational backends provide api similar to numpy, and computation can be dispatched in a compatible way.

This is a possibility to investigate, initially mainly for parts that are outside full models because those contain too much surrounding code that would need to change, e.g. base.data.

maybe that would also work for dask, e.g. #8629

2 sklearn PRs:
scikit-learn/scikit-learn#22554
scikit-learn/scikit-learn#25956

One possibility to strip base.data handling is to create a model subclass that does not call super().__init__. But it would still need to add non-data initialization like df_xxx.