试过把8卡部署一个实例，也试过4卡2实例，2卡4实例，响应都很慢，同并发下反而2卡4实例效率更高？
与之前用的910B差太远了。有没有好的优化方案呢？
启动的脚本：

services:
  deepseek_32b_1:
  # image.sourcefind.cn:5000/dcu/admin/base/custom           vllm0.8.4-ubuntu22.04-dtk25.04-rc7-das1.5-py3.10-20250429-dev-qwen3-only 
    image: 6e12a1c4ae4d
    container_name: deepseek_32b_1
    network_mode: bridge
    command: >
      vllm serve /llm/DeepSeek-R1-Distill-Qwen-32B/ --served_model_name DeepSeek-R1-Distill-Qwen-32B  --tensor-parallel-size 8 --max-model-len 32768  --enforce-eager
    ports:
      - "8468:8000"
    privileged: true
    restart: always
    devices:
      - /dev/kfd:/dev/kfd
      - /dev/dri:/dev/dri
    ipc: host
    user: root
    shm_size: 64G
    group_add:
      - video
    security_opt:
      - seccomp:unconfined
    volumes:
      - /opt/hyhal:/opt/hyhal:ro
      - /mnt/llm:/llm
    environment:
      - HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7