vLLM Metal
A community-maintained vLLM hardware plugin that runs high-performance LLM inference on Apple Silicon Macs using MLX as the compute backend.
vLLM Metal is a plugin that lets vLLM run on Apple Silicon Macs, using MLX as the primary compute backend and unifying MLX and PyTorch under a single lowering path. Upstream vLLM supplies the API server, scheduler and paged block manager, mlx_lm supplies the token-wise model layers, and vllm-metal owns the request-aware attention path: the paged variable-length-attention kernel, prefill acceleration on newer Apple silicon tensor units, and speculative decoding.
A recent release made a unified paged variable-length Metal kernel the default attention backend, reporting large gains in time-to-first-token and throughput over the prior version, and added support for running large hybrid attention models on a single Mac.
It requires macOS 15 (Sequoia) or later on Apple Silicon, and installs via Homebrew for the stable release or a curl installer for the latest development build.