Perfect for when you have unified memory and you want to use as much of it as possible for running the LLM. </sarcasm>