[dzr0001]

My token throughput is much better using vLLM-mlx on my M2 ultra than llama.cpp. It might be worth a shot to give it a try.