vLLM-mlx – 65 tok/s LLM inference on Mac with tool calling and prompt cachinggithub.com·3 pts·raullen·1