Some frameworks have tried to do this automatically, but its extremely slow. You are better off just running whatever won't fit on the CPU (which is what llama.cpp does).