快速开始¶
简介¶
本节将指导您完成基于容器的环境搭建和大模型推理,以 Qwen3-0.6B 离线单卡推理脚本为例。
- 关于使用不同模型的详细信息,请参见“模型教程”目录下的对应模型教程,例如 Qwen3-30B-A3B。
- 关于使用不同功能的详细信息,请参见“功能教程”目录下的对应功能教程,例如 Prefill-Decode 分离部署 (Deepseek)。
前提条件¶
支持的设备¶
- Atlas A2 训练系列(Atlas 800T A2、Atlas 900 A2 PoD、Atlas 200T A2 Box16、Atlas 300T A2)
- Atlas 800I A2 推理系列(Atlas 800I A2)
- Atlas A3 训练系列(Atlas 800T A3、Atlas 900 A3 SuperPoD、Atlas 9000 A3 SuperPoD)
- Atlas 800I A3 推理系列(Atlas 800I A3)
- [实验性] Atlas 300I 推理系列(Atlas 300I Duo)
环境要求¶
- 操作系统:Linux
- Python:>= 3.10, < 3.13
- 硬件:配备昇腾 NPU,通常为 Atlas 800 A2 系列。
-
软件:
软件 支持的版本 备注 Ascend HDK 参考文档 CANN 9.0.0 CANN 所需 CANN == 9.0.0 vllm-ascend 和 torch-npu 所需 torch-npu == 2.10.0 vllm-ascend 所需,无需手动安装,将在以下步骤中自动安装 torch == 2.10.0 torch-npu 和 vllm 所需,无需手动安装,将在以下步骤中自动安装 NNAL == 9.0.0 libatb.so 所需,支持高级张量运算
使用容器搭建环境¶
在使用容器之前,请确保您的系统已安装 Docker。如果尚未安装 Docker,请参考 Docker 安装指南 进行安装。
# Update DEVICE according to your device (/dev/davinci[0-7])
export DEVICE=/dev/davinci0
# Update the vllm-ascend image
# Atlas A2:
# export IMAGE=quay.io/ascend/vllm-ascend:v0.22.1rc1
# Atlas A3:
# export IMAGE=quay.io/ascend/vllm-ascend:v0.22.1rc1-a3
export IMAGE=quay.io/ascend/vllm-ascend:v0.22.1rc1
docker run --rm \
--name vllm-ascend \
--shm-size=1g \
--device $DEVICE \
--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 \
-p 8000:8000 \
-it $IMAGE bash
# Install curl
apt-get update -y && apt-get install -y curl
# Update DEVICE according to your device (/dev/davinci[0-7])
export DEVICE=/dev/davinci0
# Update the vllm-ascend image
# Atlas A2:
# export IMAGE=quay.io/ascend/vllm-ascend:v0.22.1rc1-openeuler
# Atlas A3:
# export IMAGE=quay.io/ascend/vllm-ascend:v0.22.1rc1-a3-openeuler
export IMAGE=quay.io/ascend/vllm-ascend:v0.22.1rc1-openeuler
docker run --rm \
--name vllm-ascend \
--shm-size=1g \
--device $DEVICE \
--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 \
-p 8000:8000 \
-it $IMAGE bash
# Install curl
yum update -y && yum install -y curl
默认工作目录为 /workspace,vLLM 和 vLLM Ascend 代码放置在 /vllm-workspace 中,并以开发模式(pip install -e)安装,以便开发者修改后立即生效,无需重新安装。
使用方法¶
您可以使用 ModelScope 镜像加速下载:
在昇腾 NPU 上启动 vLLM 有两种方式:
安装 vLLM 后,您可以开始为输入提示列表生成文本(即离线批量推理)。
创建并运行一个简单的推理测试。example.py 的内容可以如下:
from vllm import LLM, SamplingParams
prompts = [
"Hello, my name is",
"The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
# The first run will take about 3-5 mins (10 MB/s) to download models
llm = LLM(model="Qwen/Qwen3-0.6B")
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
然后运行:
如果遇到与 Hugging Face 的连接错误(例如 We couldn't connect to 'https://huggingface.co' to load the files, and couldn't find them in the cached files.),请运行以下命令以使用 ModelScope 作为替代:
以下部分显示 vllm 已成功检测到昇腾平台:
INFO 05-27 11:40:38 [__init__.py:44] Available plugins for group vllm.platform_plugins:
INFO 05-27 11:40:38 [__init__.py:46] - ascend -> vllm_ascend:register
INFO 05-27 11:40:38 [__init__.py:49] All plugins in this group will be loaded. Set `VLLM_PLUGINS` to control which plugins to load.
INFO 05-27 11:40:38 [__init__.py:238] Platform plugin ascend is activated
以下部分显示最终输出:
Prompt: 'Hello, my name is', Generated text: ' Lucy and I am an 8 year old who loves to draw and write stories'
Prompt: 'The president of the United States is', Generated text: " a key leader in the federal government, and the president's role in the executive"
Prompt: 'The capital of France is', Generated text: ' a city. What is the capital of France? The capital of France is Paris'
Prompt: 'The future of AI is', Generated text: ' a topic that is being discussed in various contexts. In the business world, AI'
以下部分显示离线推理后进程退出,不影响实际推理:
(EngineCore pid=970) INFO 05-12 11:36:00 [core.py:1201] Shutdown initiated (timeout=0)
(EngineCore pid=970) INFO 05-12 11:36:00 [core.py:1224] Shutdown complete
ERROR 05-12 11:36:01 [core_client.py:704] Engine core proc EngineCore died unexpectedly, shutting down client.
sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute
vLLM 也可以部署为实现了 OpenAI API 协议的服务器。运行 以下命令,使用 Qwen/Qwen3-0.6B 模型启动 vLLM 服务器:
# Deploy vLLM server (The first run will take about 3-5 mins (10 MB/s) to download models)
vllm serve Qwen/Qwen3-0.6B &
如果您看到如下日志:
INFO: Started server process [3594]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)
恭喜,您已成功启动 vLLM 服务器!
您可以查询模型列表:
您也可以使用输入提示查询模型:
curl http://localhost:8000/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen3-0.6B",
"prompt": "Beijing is a",
"max_completion_tokens": 5,
"temperature": 0
}' | python3 -m json.tool
vLLM 作为后台进程运行,您可以使用 kill -2 $VLLM_PID 优雅地停止后台进程,这类似于使用 Ctrl-C 停止前台 vLLM 进程:
输出如下:
INFO: Shutting down FastAPI HTTP server.
INFO: Shutting down
INFO: Waiting for application shutdown.
INFO: Application shutdown complete.
最后,您可以使用 ctrl-D 退出容器。