https://github.com/xiaomingx/learning-deepseek-model
使用deepseek来做大模型应用开发(how to use deepseek)
https://github.com/xiaomingx/learning-deepseek-model
awesome deepseek gpt llm openai quickstart
Last synced: 3 months ago
JSON representation
使用deepseek来做大模型应用开发(how to use deepseek)
- Host: GitHub
- URL: https://github.com/xiaomingx/learning-deepseek-model
- Owner: XiaomingX
- License: apache-2.0
- Created: 2024-12-17T04:08:19.000Z (about 1 year ago)
- Default Branch: main
- Last Pushed: 2025-02-06T02:15:55.000Z (12 months ago)
- Last Synced: 2025-04-01T04:38:51.843Z (10 months ago)
- Topics: awesome, deepseek, gpt, llm, openai, quickstart
- Language: Python
- Homepage: https://x.com/seclink
- Size: 9.77 KB
- Stars: 17
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# use-deepseek-py
🔥 使用 DeepSeek 构建大模型应用的 Python 工具包
## 🚀 快速开始
## 📊 模型选型指南
### 性能对比表
| 模型名称 | 参数量 | 上下文长度 | 适用场景 | 推理速度 (tokens/s) |
|--------------------------|--------|------------|-------------------|---------------------|
| deepseek-v3-base | 671B | 128k | 代码生成/数据分析 | 320 (A100) |
| deepseek-r1-32b-distill | 32B | 64k | 数学推理/逻辑推导 | 580 (RTX 4090) |
| deepseek-r1-zero | 37B | 128k | 多轮对话/创意写作 | 420 (A100) |
💡 建议搭配 vLLM 推理框架实现吞吐量优化
## 🛠️ 进阶功能
### API 对接
```python
import os
from deepseek import DeepSeekAPI
client = DeepSeekAPI(
api_key=os.getenv("DEEPSEEK_KEY"),
base_url="https://api.deepseek.com/v1"
)
# 带思维链的复杂推理
response = client.chat.completions.create(
model="deepseek-reasoner",
messages=[{"role": "user", "content": "证明勾股定理"}],
temperature=0.7,
reasoning_depth=2 # 启用二级推理验证
)
print(response.choices[0].message.reasoning_content) # 显示验证过程
```
### 本地部署优化
```yaml
# config.yaml
compute:
precision: bfloat16
parallelism:
tensor: 2
pipeline: 4
optimization:
flash_attention: true
continuous_batching: true
quantization:
enabled: true
method: awq
```
## 🌟 核心优势
### 技术创新
- **动态专家选择**：仅激活 37B/671B 参数，降低 95% 计算成本
- **多模态注意力**：MLA 机制提升长文本理解能力（128k tokens）
- **强化学习优化**：GRPO 算法增强推理能力
### 开源生态
```mermaid
graph TD
A[DeepSeek 模型] --> B(代码生成)
A --> C(数据分析)
A --> D(数学证明)
B --> E[VSCode 插件]
C --> F[Pandas 扩展]
D --> G[Lean4 集成]
```
## 🏆 性能基准
| 测试集 | DeepSeek-R1 | GPT-4 | 提升幅度 |
|----------------|-------------|-------|---------|
| HumanEval | 73.78% | 67% | +10% |
| GSM8K | 84.1% | 80% | +5% |
| MATH-500 | 68.3% | 65% | +5% |
| 训练能耗 (PFLOPs)| 2.8M | 12M | -76% |
数据来源：DeepSeek 技术白皮书
## 🛡️ 最佳实践
1. **成本控制**
```python
# 启用动态批处理
model.set_optimization(
max_batch_size=32,
memory_utilization=0.85
)
```
2. **错误处理**
```python
try:
response = model.generate(...)
except DeepSeekError as e:
if "rate_limit" in str(e):
print("触发限流，自动重试中...")
time.sleep(1)
response = model.generate(...)
```
3. **性能监控**
```bash
deepseek-monitor --model deepseek-r1 --metrics latency throughput error_rate
```
*本项目的模型权重遵循 DeepSeek 社区许可协议，商业使用需遵守[附加条款](https://www.deepseek.com/license)*