DeepSeek R1 简易指南：架构、本地部署和硬件要求

import openai
client = openai.OpenAI(
    base_url="https://api.deepseek.com/v1",
    api_key="your_api_key"
)
response = client.chat.completions.create(
    model="deepseek-r1",
    messages=[{"role":"user","content":"解释量子纠缠现象"}]
)

# 完整API接入示例（Python）
import openai
from dotenv import load_dotenv
import os

# 环境配置
load_dotenv()
DEEPSEEK_API_KEY = os.getenv("DEEPSEEK_API_KEY")

# 客户端初始化
client = openai.OpenAI(
    base_url="https://api.deepseek.com/v1",
    api_key=DEEPSEEK_API_KEY,
    timeout=30  # 超时设置
)

# 带重试机制的请求函数
def query_deepseek(prompt, max_retries=3):
    for attempt in range(max_retries):
        try:
            response = client.chat.completions.create(
                model="deepseek-r1",
                messages=[{"role": "user", "content": prompt}],
                temperature=0.7,
                top_p=0.9,
                max_tokens=1024
            )
            return response.choices[0].message.content
        except Exception as e:
            if attempt == max_retries - 1:
                raise e
            print(f"请求失败，正在重试... ({attempt+1}/{max_retries})")

# 使用示例
if __name__ == "__main__":
    result = query_deepseek("用React实现可拖拽的甘特图组件")
    print(result)

| 模型类型        | 最小GPU配置      | CPU配置          | 内存要求  | 磁盘空间 |
|---------------|----------------|------------------|---------|--------|
| R1-Zero全量版  | RTX 4090(24GB) | Xeon 8核+128GB   | 128GB   | 500GB  |
| R1蒸馏版-70B   | RTX 3090(24GB) | i9-13900K+64GB   | 64GB    | 320GB  |
| R1蒸馏版-14B   | RTX 3060(12GB) | Ryzen 7+32GB     | 32GB    | 80GB   |
| R1蒸馏版-1.5B  | 无需GPU         | 任意四核处理器+8GB | 8GB     | 12GB   |

# 完整部署流程（Ubuntu示例）
# 步骤1：安装依赖
sudo apt update && sudo apt install -y nvidia-driver-535 cuda-12.2

# 步骤2：安装Ollama
curl -fsSL https://ollama.com/install.sh | sh

# 步骤3：配置环境变量
echo 'export OLLAMA_HOST=0.0.0.0' >> ~/.bashrc
source ~/.bashrc

# 步骤4：启动服务
sudo systemctl start ollama

# 步骤5：拉取模型（以14B为例）
ollama pull deepseek-r1:14b

# 步骤6：运行模型（带GPU加速）
ollama run deepseek-r1:14b --gpu

# 步骤7：验证部署
curl http://localhost:11434/api/tags | jq

方案一：vLLM服务化部署
# 启动推理服务
vllm serve --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B \
--tensor-parallel-size 2 \
--max-model-len 32768 \
--gpu-memory-utilization 0.9

# 客户端调用
from vllm import LLM, SamplingParams
llm = LLM("deepseek-ai/DeepSeek-R1-Distill-Qwen-32B")
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
print(llm.generate(["解释BERT模型的注意力机制"], sampling_params))

方案二：llama.cpp量化部署
# 模型转换
./quantize ./models/deepseek-r1-14b.gguf ./models/deepseek-r1-14b-Q5_K_M.gguf Q5_K_M

# 启动推理
./main -m ./models/deepseek-r1-14b-Q5_K_M.gguf \
-n 1024 \
--repeat_penalty 1.1 \
--color \
-i

# docker-compose.yml配置示例
version: '3.8'

services:
  ollama:
    image: ollama/ollama
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ollama:/root/.ollama
    ports:
      - "11434:11434"

  api-gateway:
    image: nginx:alpine
    ports:
      - "80:80"
    volumes:
      - ./nginx.conf:/etc/nginx/nginx.conf

volumes:
  ollama: