AI模型从开发到部署，Docker已经成为事实标准。但AI应用的容器化比普通Web应用复杂得多——GPU支持、大模型文件、内存管理、推理优化……本文将手把手带你从零搭建一个生产级的AI应用Docker部署方案。

为什么AI应用需要Docker？

基础：Dockerfile编写

1. 选择基础镜像

# 轻量级（仅CPU推理）
FROM python:3.11-slim

# GPU支持（推荐）
FROM nvidia/cuda:12.4.1-runtime-ubuntu22.04

# 完整开发环境
FROM nvidia/cuda:12.4.1-devel-ubuntu22.04

2. 完整的AI应用Dockerfile

FROM nvidia/cuda:12.4.1-runtime-ubuntu22.04

# 设置环境变量
ENV DEBIAN_FRONTEND=noninteractive
ENV PYTHONUNBUFFERED=1
ENV CUDA_HOME=/usr/local/cuda
ENV PATH=${CUDA_HOME}/bin:${PATH}
ENV LD_LIBRARY_PATH=${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}

# 安装系统依赖
RUN apt-get update && apt-get install -y --no-install-recommends \
    python3.11 \
    python3-pip \
    python3.11-venv \
    libgl1-mesa-glx \
    libglib2.0-0 \
    && rm -rf /var/lib/apt/lists/*

# 创建工作目录
WORKDIR /app

# 安装Python依赖（分层缓存优化）
COPY requirements.txt .
RUN pip3 install --no-cache-dir -r requirements.txt

# 复制应用代码
COPY . .

# 下载模型（构建时下载，避免运行时下载）
RUN python3 download_model.py

# 暴露端口
EXPOSE 8000

# 健康检查
HEALTHCHECK --interval=30s --timeout=10s --retries=3 \
    CMD curl -f http://localhost:8000/health || exit 1

# 启动命令
CMD ["python3", "app.py"]

GPU支持配置

Docker Compose配置

version: '3.8'

services:
  ai-app:
    build: .
    ports:
      - "8000:8000"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    environment:
      - CUDA_VISIBLE_DEVICES=0
      - MODEL_PATH=/app/models
    volumes:
      - ./models:/app/models
      - ./logs:/app/logs
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
      interval: 30s
      timeout: 10s
      retries: 3

多GPU配置

deploy:
  resources:
    reservations:
      devices:
        - driver: nvidia
          device_ids: ['0', '1']  # 指定GPU
          capabilities: [gpu]

实战：部署一个LLM推理服务

应用代码 (app.py)

from fastapi import FastAPI
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

app = FastAPI(title="AI Inference Service")

# 加载模型
model_name = "Qwen/Qwen2.5-7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

@app.get("/health")
async def health():
    return {"status": "healthy"}

@app.post("/generate")
async def generate(prompt: str, max_tokens: int = 512):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_tokens,
            temperature=0.7,
            top_p=0.9
        )
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return {"response": response}

requirements.txt

fastapi==0.115.0
uvicorn==0.30.0
transformers==4.44.0
torch==2.4.0
accelerate==0.33.0

生产环境优化

1. 多阶段构建减小镜像体积

# 构建阶段
FROM nvidia/cuda:12.4.1-devel-ubuntu22.04 AS builder
WORKDIR /build
COPY requirements.txt .
RUN pip3 install --user --no-cache-dir -r requirements.txt

# 运行阶段
FROM nvidia/cuda:12.4.1-runtime-ubuntu22.04
COPY --from=builder /root/.local /root/.local
ENV PATH=/root/.local/bin:$PATH
WORKDIR /app
COPY . .
CMD ["python3", "app.py"]

2. 使用vLLM加速推理

# 使用vLLM官方镜像
FROM vllm/vllm-openai:latest

COPY app.py /app/app.py
WORKDIR /app

CMD ["--model", "Qwen/Qwen2.5-7B-Instruct", \
     "--host", "0.0.0.0", \
     "--port", "8000", \
     "--tensor-parallel-size", "1"]

3. 模型文件优化

# 使用GGUF格式（CPU友好）
python3 convert_to_gguf.py --model ./model --outtype q4_k_m

# 使用量化减少内存占用
# 7B模型：FP16需14GB，INT4仅需4GB

监控与日志

Prometheus监控

from prometheus_fastapi_instrumentator import Instrumentator

app = FastAPI()
Instrumentator().instrument(app).expose(app)

# 自定义指标
from prometheus_client import Counter, Histogram

REQUEST_COUNT = Counter('inference_requests_total', 'Total inference requests')
INFERENCE_TIME = Histogram('inference_duration_seconds', 'Inference duration')

@app.post("/generate")
async def generate(prompt: str):
    REQUEST_COUNT.inc()
    with INFERENCE_TIME.time():
        # 推理逻辑
        pass

常见问题与解决方案

总结

Docker部署AI应用的关键在于：选择合适的基础镜像、正确配置GPU支持、优化模型加载和推理性能。掌握这些技巧后，你可以将任何AI应用快速部署到生产环境。

问题	原因	解决方案
CUDA out of memory	模型太大	使用量化、减小batch_size
容器启动慢	模型加载时间长	预热脚本、模型缓存
GPU不可用	驱动不匹配	统一host和container的CUDA版本
推理速度慢	CPU瓶颈	使用vLLM、TensorRT