1、CPU：Intel(R) Xeon(R) Silver 4410Y 2、DCU：K100\_AI \* 4 3、镜像：image.sourcefind.cn:5000/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.1-rc5-rocblas101839-0811-das1.6-py3.10-20250913-rc1 4、启动命令： ``` export W4A16_MOE_CUDA=0; vllm serve --model=/mnt/model/Qwen3-235B-A22B-Instruct-2507-AWQ \ --quantization moe_wna16 \ --dtype float16 \ --tensor-parallel-size 4 \ --served-model-name Qwen3-235B-A22B-AWQ \ --max-model-len 32768 \ --compilation-config '{"full_cuda_graph": true}' \ --gpu_memory_utilization 0.92 \ --port 28000 ``` 5、启动成功后，测试 token 单并发只能达到 15t/s，且 K100_ai 显卡利用率不高，请问有什么办法能够提高 token 速率？ ``` /opt/hyhal/bin/hy-smi ============================ System Management Interface ============================= ====================================================================================== DCU Temp AvgPwr Perf PwrCap VRAM% DCU% Mode 0 53.0C 180.0W auto 400.0W 94% 24.2% Normal 1 54.0C 182.0W auto 400.0W 94% 12.5% Normal 2 54.0C 183.0W auto 400.0W 94% 24.2% Normal 3 53.0C 187.0W auto 400.0W 94% 6.7% Normal ```