vllm serve PowerInfer/SmallThinker-3B-Preview --trust-remote-code --dtype bfloat16 --max-seq-len-to-capture 32768 -tp 1 --max-model-len 32768
