ggml and pytorch support mps. aside from bleeding edge most ml workloads can run on mps these days.