You wouldnt be using coreml since thats apples software, but likely vulkan versions of llama.cpp or pytorch or what have you. How well that works at the moment i dont know

To my understanding the M1 and M2 have very functional GPU drivers. I don't know about the other specifics you asked