The python overhead of launching big ML jobs is nontrivial, so I think speeding that up would be meaningful. (I mean the initial tracing and other setup, not things once the GPUs are actually doing the work).

Also you can use projects like numba https://numba.pydata.org/