OpenAI 兼容 API 服务器对比：vLLM/TGI/Ollama/LM Studio

┌─────────────────────────────────────────────┐
│         任意 OpenAI SDK 客户端               │
│   (Python / TypeScript / Go / Rust ...)     │
├─────────────────────────────────────────────┤
│         OpenAI 兼容 API 层                   │
│   /v1/chat/completions                      │
│   /v1/completions                           │
│   /v1/embeddings                            │
│   /v1/models                                │
├─────────────┬───────────┬───────────────────┤
│   vLLM      │   TGI     │   Ollama          │
│             │           │   LM Studio       │
└─────────────┴───────────┴───────────────────┘

from openai import OpenAI

# 同一份代码，切换后端只需改一行
# OpenAI: client = OpenAI(base_url="https://api.openai.com/v1", api_key="sk-xxx")
# vLLM:   client = OpenAI(base_url="http://vllm:8000/v1", api_key="none")
# Ollama: client = OpenAI(base_url="http://ollama:11434/v1", api_key="ollama")
# TGI:    client = OpenAI(base_url="http://tgi:8080/v1", api_key="none")

response = client.chat.completions.create(
    model="meta-llama/Llama-3.1-8B-Instruct",
    messages=[{"role": "user", "content": "Hello"}],
    stream=True
)

response = client.chat.completions.create(
    model="llama3.1:8b",
    messages=[{"role": "user", "content": "What's the weather in Tokyo?"}],
    tools=[{
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current weather",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string"}
                },
                "required": ["city"]
            }
        }
    }],
    tool_choice="auto"
)

# 决策树
def choose_engine(scenario):
    if scenario == "开发原型":
        return "Ollama"  # 一键安装，最简单
    elif scenario == "桌面应用":
        return "LM Studio"  # GUI 界面，非技术用户友好
    elif scenario == "高并发API":
        return "vLLM"  # 吞吐最高，PagedAttention
    elif scenario == "HuggingFace生态":
        return "TGI"  # 模型兼容性最好
    elif scenario == "多轮对话/Agent":
        return "SGLang"  # RadixAttention（额外选项）
    elif scenario == "CPU推理":
        return "Ollama + llama.cpp"  # GGUF 量化
    elif scenario == "边缘设备":
        return "llama.cpp"  # 最轻量
    elif scenario == "企业生产":
        return "vLLM + Open WebUI"  # 最佳组合
    else:
        return "vLLM"  # 默认选择

# docker-compose.multi-engine.yml
version: "3.9"
services:
  vllm:
    image: vllm/vllm-openai:latest
    ports: ["8000:8000"]
    command:
      - --model=meta-llama/Llama-3.1-70B-Instruct
      - --tensor-parallel-size=4
      - --enable-prefix-caching
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 4
              capabilities: [gpu]

  ollama:
    image: ollama/ollama:latest
    ports: ["11434:11434"]
    volumes:
      - ollama_data:/root/.ollama
    environment:
      - OLLAMA_NUM_PARALLEL=4

  nginx:
    image: nginx:alpine
    ports: ["443:443"]
    volumes:
      - ./nginx.conf:/etc/nginx/nginx.conf
    depends_on: [vllm, ollama]

volumes:
  ollama_data:

# nginx.conf - 按 model 名称路由到不同引擎
upstream vllm_backend {
    server vllm:8000;
}
upstream ollama_backend {
    server ollama:11434;
}

server {
    listen 443 ssl;
    server_name llm.internal;

    location /v1/chat/completions {
        # 读取请求体中的 model 字段决定路由
        # 70B → vLLM, 8B → Ollama
        access_by_lua_block {
            ngx.req.read_body()
            local body = ngx.req.get_body_data()
            if body and string.find(body, "70B") then
                ngx.var.upstream = "vllm_backend"
            else
                ngx.var.upstream = "ollama_backend"
            end
        }
        proxy_pass http://$upstream;
    }
}