TecGPT / docs /use_vllm.md
Gnib's picture
Upload 834 files
444f09e verified
|
raw
history blame
No virus
1.27 kB
# 使用VLLM
## 1. 首先启动 VLLM,自行选择模型
```
python -m vllm.entrypoints.openai.api_server --model /home/hmp/llm/cache/Qwen1___5-32B-Chat --tensor-parallel-size 2 --dtype=half
```
这里使用了存储在 `/home/hmp/llm/cache/Qwen1___5-32B-Chat` 的本地模型,可以根据自己的需求更改。
## 2. 测试 VLLM
```
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "/home/hmp/llm/cache/Qwen1___5-32B-Chat",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "怎么实现一个去中心化的控制器?"}
]
}'
```
## 3. 配置本项目
```
API_KEY = "sk-123456789xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx123456789"
LLM_MODEL = "vllm-/home/hmp/llm/cache/Qwen1___5-32B-Chat(max_token=4096)"
API_URL_REDIRECT = {"https://api.openai.com/v1/chat/completions": "http://localhost:8000/v1/chat/completions"}
```
```
"vllm-/home/hmp/llm/cache/Qwen1___5-32B-Chat(max_token=4096)"
其中
"vllm-" 是前缀(必要)
"/home/hmp/llm/cache/Qwen1___5-32B-Chat" 是模型名(必要)
"(max_token=6666)" 是配置(非必要)
```
## 4. 启动!
```
python main.py
```