AMD显卡运行qwen3.5-使用VLLM

1、下载镜像

docker pull docker.1ms.run/vllm/vllm-openai-rocm:v0.18.0

2、自定义docker镜像

FROM docker.1ms.run/vllm/vllm-openai-rocm:v0.18.0
RUN pip install modelscope>=1.18.1

docker build -t my-vllm-with-modelscope:v0.18.0 .

3、运行

docker run --rm \
  --name vllm_qwen35 \
  --device=/dev/kfd --device=/dev/dri \
  --security-opt seccomp=unconfined \
  --group-add video \
  --ipc=host \
  -p 8000:8000 \
  -v /home/vllm:/root/.cache/modelscope \
  -e VLLM_USE_MODELSCOPE=True \
  my-vllm-with-modelscope:nightly  \
  --model tclf90/Qwen3.5-9B-AWQ \
  --language-model-only \
  --tensor-parallel-size 1 \
  --reasoning-parser qwen3 \
  --enable-prefix-caching \
 --max-model-len 32768