docker-compose部署Qwen3-Next-80B-A3B-Instruct大模型的步骤

安装docker环境

 apt  install [docker.io](http://docker.io/)

安装docker-comopse环境

sudo ln -s /root/workspace/docker-compose-linux-x86_64 /usr/local/bin/docker-compose

安装NVIDIA Container Toolkit

curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey \
 | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg

curl -fsSL https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list \
 | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#' \
 | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list

sudo apt-get update

sudo apt-get install -y nvidia-container-toolkit

sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

docker run --rm --gpus all nvidia/cuda:12.4.1-base-ubuntu22.04 nvidia-smi

下载models：Qwen3-Next-80B-A3B-Instruct

└── qwen3
    ├── LICENSE
    ├── README.md
    ├── config.json
    ├── generation_config.json
    ├── merges.txt
    ├── model-00001-of-00041.safetensors
    ├── model-00002-of-00041.safetensors
    ├── model-00003-of-00041.safetensors
    ├── model-00004-of-00041.safetensors
    ├── model-00005-of-00041.safetensors
    ├── model-00006-of-00041.safetensors
    ├── model-00007-of-00041.safetensors
    ├── model-00008-of-00041.safetensors
    ├── model-00009-of-00041.safetensors
    ├── model-00010-of-00041.safetensors
    ├── model-00011-of-00041.safetensors
    ├── model-00012-of-00041.safetensors
    ├── model-00013-of-00041.safetensors
    ├── model-00014-of-00041.safetensors
    ├── model-00015-of-00041.safetensors
    ├── model-00016-of-00041.safetensors
    ├── model-00017-of-00041.safetensors
    ├── model-00018-of-00041.safetensors
    ├── model-00019-of-00041.safetensors
    ├── model-00020-of-00041.safetensors
    ├── model-00021-of-00041.safetensors
    ├── model-00022-of-00041.safetensors
    ├── model-00023-of-00041.safetensors
    ├── model-00024-of-00041.safetensors
    ├── model-00025-of-00041.safetensors
    ├── model-00026-of-00041.safetensors
    ├── model-00027-of-00041.safetensors
    ├── model-00028-of-00041.safetensors
    ├── model-00029-of-00041.safetensors
    ├── model-00030-of-00041.safetensors
    ├── model-00031-of-00041.safetensors
    ├── model-00032-of-00041.safetensors
    ├── model-00033-of-00041.safetensors
    ├── model-00034-of-00041.safetensors
    ├── model-00035-of-00041.safetensors
    ├── model-00036-of-00041.safetensors
    ├── model-00037-of-00041.safetensors
    ├── model-00038-of-00041.safetensors
    ├── model-00039-of-00041.safetensors
    ├── model-00040-of-00041.safetensors
    ├── model-00041-of-00041.safetensors
    ├── model.safetensors.index.json
    ├── tokenizer.json
    ├── tokenizer_config.json
    └── vocab.json

拷贝模型服务到服务器

cd Qwen80B_LLM_Server

docker-compose up -d --build

docker ps

root@iZ6webx89mskfrwmhgi2btZ:~/workspace/Qwen80B_LLM_Server# docker ps
CONTAINER ID   IMAGE                    COMMAND                  CREATED          STATUS                    PORTS                                                                      NAMES
b08a84029791   nginx:alpine             "/docker-entrypoint.…"   17 minutes ago   Up 17 minutes             0.0.0.0:80->80/tcp, :::80->80/tcp, 0.0.0.0:443->443/tcp, :::443->443/tcp   vllm-nginx
108eea5174d9   vllm-qwen3-vllm-server   "python3 -m vllm.ent…"   17 minutes ago   Up 17 minutes (healthy)   0.0.0.0:8000->8000/tcp, :::8000->8000/tcp                                  qwen3-vllm-server

Qwen80B_LLM_Server的主要代码

DockerFile

# vLLM with Updated Transformers - 支持Qwen3架构
FROM vllm/vllm-openai:latest

# 设置工作目录
WORKDIR /workspace

# 更新包管理器
RUN apt-get update && apt-get install -y \
    curl \
    git \
    && rm -rf /var/lib/apt/lists/*

# 升级Python包管理工具
RUN pip install --no-cache-dir --upgrade pip setuptools wheel

# 升级transformers和相关依赖到最新版本
RUN pip install --no-cache-dir --upgrade \
    transformers>=4.45.0 \
    tokenizers>=0.19.0 \
    accelerate>=0.34.0 \
    torch>=2.1.0 \
    torchvision \
    torchaudio

# 从源码安装最新transformers（支持qwen3_next）
RUN pip install --no-cache-dir --force-reinstall \
    git+https://github.com/huggingface/transformers.git

# 重新安装vLLM以确保兼容性
RUN pip install --no-cache-dir --upgrade --force-reinstall vllm

# 安装其他可能需要的依赖
RUN pip install --no-cache-dir \
    huggingface-hub \
    safetensors \
    numpy \
    scipy

# 验证安装
RUN python3 -c "import transformers; print(f'Transformers version: {transformers.__version__}')" && \
    python3 -c "import vllm; print(f'vLLM version: {vllm.__version__}')" && \
    python3 -c "from transformers import AutoConfig; print('Transformers import successful')"

# 设置环境变量
ENV PYTHONPATH=/workspace
ENV HF_HUB_ENABLE_HF_TRANSFER=1
ENV CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

# 创建日志目录
RUN mkdir -p /workspace/logs

# 健康检查脚本
COPY <<EOF /workspace/healthcheck.sh
#!/bin/bash
curl -f http://localhost:8000/health || exit 1
EOF

RUN chmod +x /workspace/healthcheck.sh

# 暴露端口
EXPOSE 8000

# 默认启动命令（可被docker-compose覆盖）
CMD ["python3", "-m", "vllm.entrypoints.openai.api_server", "--help"]

.env文件

# VLLM服务配置
VLLM_API_KEY=XXXXXXX
MODEL_PATH=/root/workspace/models
HOST_PORT=8000

# GPU配置
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

# 其他配置
COMPOSE_PROJECT_NAME=vllm-qwen3

docker-compose.yml

services:
  vllm-server:
    build:
      context: .
      dockerfile: Dockerfile
    # 如果已经构建了镜像，可以直接使用：
    # image: vllm-qwen3:latest
    container_name: qwen3-vllm-server
    restart: unless-stopped
    
    # GPU配置
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    
    # 环境变量
    environment:
      - VLLM_API_KEY=${VLLM_API_KEY}
      - CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
      - PYTHONPATH=/workspace
      - HF_HUB_ENABLE_HF_TRANSFER=1
      - TRANSFORMERS_CACHE=/root/.cache/transformers
    
    # 端口映射
    ports:
      - "${HOST_PORT:-8000}:8000"
    
    # 卷挂载
    volumes:
      - ${MODEL_PATH}:/workspace/models:ro
      - vllm_cache:/root/.cache
      - ./logs:/workspace/logs
    
    # 启动命令
    command: >
      --model /workspace/models/qwen3
      --served-model-name Qwen3-Next-80B-A3B-Instruct
      --host 0.0.0.0
      --port 8000
      --tensor-parallel-size 8
      --gpu-memory-utilization 0.88
      --max-model-len 32768
      --max-num-seqs 256
      --max-num-batched-tokens 32768
      --trust-remote-code
      --disable-log-requests
      --dtype bfloat16
      --swap-space 32
      --enable-chunked-prefill
      --disable-custom-all-reduce
      --block-size 32
      --api-key ${VLLM_API_KEY}
    
    # 健康检查 - 延长启动时间
    healthcheck:
      test: ["CMD", "/workspace/healthcheck.sh"]
      interval: 30s
      timeout: 15s
      retries: 5
      start_period: 900s  # 15分钟启动时间，因为需要升级依赖
    
    # 网络配置
    networks:
      - vllm-network
      
volumes:
  vllm_cache:
    driver: local

networks:
  vllm-network:
    driver: bridge

测试服务启动是否正常

from openai import OpenAI
import time
import traceback

def run80B(prompt: str):
    print("use run80B try")
    client = OpenAI(
        api_key="shgbitai@2025",  # 修正了这里的语法
        base_url="http://8.216.37.224:8000/v1"
    )

    reasoning_content = ""
    answer_content = ""
    is_answering = False

    try:
        astart_time = time.perf_counter()
        completion = client.chat.completions.create(
            model="Qwen3-Next-80B-A3B-Instruct",
            messages=[
                {"role": "user", "content": prompt}
            ],
            stream=True,
        )
    except Exception as e:
        traceback.print_exc()
        answer_content = "API调用失败，请检查服务状态"
        return answer_content

    for chunk in completion:
        if not chunk.choices:
            if hasattr(chunk, 'usage'):
                print("\nUsage:")
                print(chunk.usage)
        else:
            delta = chunk.choices[0].delta

            # 处理推理内容
            if hasattr(delta, 'reasoning_content') and delta.reasoning_content is not None:
                reasoning_content += delta.reasoning_content
            else:
                # 处理回答内容
                if delta.content and not is_answering:
                    is_answering = True

                if delta.content:
                    answer_content += delta.content

    aend_time = time.perf_counter()
    api_time = aend_time - astart_time
    print(f"大模型调用完成，耗时: {api_time:.2f}秒")
    print(f"输出: {answer_content}")
    return answer_content

if __name__ == '__main__':
    result = run80B("你是什么模型")
    print("最终结果:", result)