目前用的镜像：image.sourcefind.cn:5000/dcu/admin/base/vllm:0.8.5-ubuntu22.04-dtk25.04.1-rc5-das1.6-py3.10-20250711 docker-compose参数： ``` services: qwen2.5-vl-72b: container_name: qwen2.5-vl-72b image: image.sourcefind.cn:5000/dcu/admin/base/vllm:0.8.5-ubuntu22.04-dtk25.04.1-rc5-das1.6-py3.10-20250711 network_mode: "host" volumes: - /dev/kfd:/dev/kfd - /dev/mkfd:/dev/mkfd - /dev/dri:/dev/dri - /opt/hyhal:/opt/hyhal - /ai:/ai # restart: always privileged: true shm_size: 64gb group_add: - "video" - "render" - "39" cap_add: - "SYS_PTRACE" security_opt: - "seccomp=unconfined" environment: - VLLM_USE_V1=0 - ALLREDUCE_STREAM_WITH_COMPUTE=1 - HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 - TZ=Asia/Shanghai command: [ "vllm", "serve", "/ai/models/Qwen2.5-VL-72B-Instruct", "--served-model-name", "Qwen2.5-VL-72B-Instruct", "--gpu-memory-utilization", "0.98", "--tensor_parallel_size", "8", "--dtype", "float16", "--trust-remote-code", "--enforce-eager", # "--no-enable-chunked-prefill", "--disable-custom-all-reduce", "--max-model-len", "32768", # "--limit-mm-per-prompt.video", "0" "--max_num_seqs", "50"] ports: - "8000:8000" ``` 报错： qwen2.5-vl-72b | loc("/usr/local/lib/python3.10/dist-packages/flash_attn/ops/triton/rotary.py":12:0): error: unsupported target: 'gfx906'