> ## Content Index
> Fetch the complete content index at: https://lucent.blog/llms.txt
> Use this file to discover other available public pages before exploring further.

# 使用vllm-ascend在昇腾910B部署千问3(Qwen3)
- URL: https://lucent.blog/7hju778q/
- Published: 2025-05-02T07:06:00.000Z
- Updated: 2026-08-21T00:49:57.000Z
- Author: Lucent
- Tags: 大模型, AI

千问3已经发布几天了，华为公开最新的Mindeie镜像，我们可以暂时使用vllm-ascend进行推理

# 下载镜像

```
docker pull quay.io/ascend/vllm-ascend:v0.8.4rc2
```

# 下载模型

我们这里以Qwen3-32B为例，其它模型同理

权重在这里下载: [https://www.modelscope.cn/models/Qwen/Qwen3-32B](https://www.modelscope.cn/models/Qwen/Qwen3-32B?ref=lucent.blog)

下载到你想要的位置

# 编写启动命令

```
docker run --rm -d \
--name Qwen3-32B \
--device /dev/davinci4 \
--device /dev/davinci5 \
--device /dev/davinci6 \
--device /dev/davinci7 \
--device /dev/davinci_manager \
--device /dev/devmm_svm \
--device /dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /root/.cache:/root/.cache \
-v /data/4pd-workspace/models/Qwen3-32B:/data1/Qwen3-32B:ro \
-p 18034:8000 \
-e PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256 \
quay.io/ascend/vllm-ascend:v0.8.4rc2 \
vllm serve /data1/Qwen3-32B --max_model_len 32768 -tp 4 --served-model-name Qwen3-32B
```

解释

- vllm serve /data1/Qwen3-32B 启动推理的命令，后面是模型权重路径
- \--max\_model\_len 32768 指定上下文长度32k
- \-tp 4 指定4卡推理，其实2卡就够，看你需要
- \--served-model-name Qwen3-32B 指定推理服务的模型名称
- \-v /data/4pd-workspace/models/Qwen3-32B:/data1/Qwen3-32B:ro 映射模型权重路径
- \-p 18034:8000 映射端口，vllm内部默认使用了8000端口推理

# 启动

执行上面的命令后，我们可以查看容器状态

```
docker ps |grep Qwen3-32B
```

![](https://img-1251540275.cos.ap-shanghai.myqcloud.com/blog/image-wkxgpnyn.png)

# 测试

```
curl "http://localhost:18034/v1/chat/completions" -H "Content-Type: application/json" \
    -H "Authorization: Bearer xxx" \
    -d '{
        "model": "Qwen3-32B",
        "stream": true,
        "messages": [
            {
                "role": "user",
                "content": "/no_think 你好."
            }
        ]
    }'
```

![](https://img-1251540275.cos.ap-shanghai.myqcloud.com/blog/image-wouykvnu.png)

至此，使用vllm-ascend在昇腾910B上部署千问3已经成功