Single request depth on vllm with dspark, I'm getting ~200 tps, I'd say it's pretty snappy.