vllm serve model \
    --tensor-parallel-size 4