I was wondering if it would make sense to decouple pytorch’s python frontend and C++ core in the same way like jax and jaxlib. Since most of the Pytorch 2.0 development is focused around python (Dynamo’s symbolic interpreter, fx tracing and pattern matching, and inductor) it would be easier to focus on python side of the development without building C++ heavy core, aten and c10 libraries just for python development.
Thanks for the suggestion.
I don’t think there is any plan to go towards that today no. Even the new pt2 stack relies very heavily on C++ code and components to achieve the appropriate speed.