python3 -m infera.engine.sglang --model-path /data/yihou_hf_cache/models/Qwen3.6-27B --served-model-name Qwen3.6-27B --tp-size 1 --trust-remote-code --host 127.0.0.1 --port 18122 --context-length 32768 --mem-fraction-static 0.85 --watchdog-timeout 3600 --enable-metrics --reasoning-parser qwen3 --cuda-graph-backend-decode full --cuda-graph-backend-prefill disabled --cuda-graph-max-bs-decode 32 --advertise-host 127.0.0.1 --etcd-endpoint 127.0.0.1:18123 --discovery-backend etcd --request-transport http --kv-event-transport zmq --kv-events-bind tcp://127.0.0.1:18125 --kv-snapshot-host 127.0.0.1 --kv-snapshot-port 18126 --enable-cache-report
