docker pull quay.io/ascend/vllm-ascend:v0.15.0rc1

docker run -it --name Qwen-dev \
> --privileged \
> --network=host \
> --shm-size=600G -w=/dev/shm/ \
> --device=/dev/davinci_manager \
> --device=/dev/devmm_svm \
> --device=/dev/hisi_hdc \
> -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
> -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
> -v /var/log/:/var/log \
> -v /usr/local/Ascend/add-ons/:/usr/local/Ascend/add-ons/ \
> -v /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi \
> -v /dev/shm/:/dev/shm/ \
> -v /usr/local/dcmi:/usr/local/dcmi -v /dev:/dev \
> -v /data/:/data/ \
> -v /mnt/:/mnt/ \
> -e ASCEND_RT_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" \
> quay.io/ascend/vllm-ascend:v0.15.0rc1 \
> /bin/bash

docker exec -it Qwen-dev bash

#创建下载脚本
vim download.py
"""
from modelscope import snapshot_download
model_path="Qwen/Qwen3-30B-A3B"
cache_path="/data/models/"
snapshot_download(model_path, cache_dir=cache_path)
"""

#下载依赖库
pip install modelscope

#运行脚本
python download.py

# 设置 max_split_size_mb来减少内存碎片并避免内存溢出（OOM）
export PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256

vllm serve /data/models/Qwen/Qwen3-30B-A3B --tensor-parallel-size 4 --enable-expert-parallel

curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "/data/models/Qwen/Qwen3-30B-A3B",
  "messages": [
    {"role": "user", "content": "Give me a short introduction to large language models."}
  ],
  "temperature": 0.6,
  "top_p": 0.95,
  "top_k": 20,
  "max_completion_tokens": 4096
}'