[大模型]DeepSeek-MoE-16b-chat FastApi 部署调用

# 开启学术加速以提升GitHub访问速度
source /etc/network_turbo
<h1>升级pip</h1><p>python -m pip install --upgrade pip</p><h1>更换为清华源加速包安装</h1><p>pip config set global.index-url <a href="https://www.php.cn/link/a6455ffc4e47fd737db213366771ec0e">https://www.php.cn/link/a6455ffc4e47fd737db213366771ec0e</a></p><h1>安装核心库</h1><p>pip install modelscope transformers sentencepiece accelerate fastapi uvicorn requests streamlit transformers_stream_generator</p><h1>安装Flash Attention加速组件（可选）</h1><p>pip install <a href="https://www.php.cn/link/0e7adb08b43a589df528d2bdd69b6b03">https://www.php.cn/link/0e7adb08b43a589df528d2bdd69b6b03</a>

import torch
from modelscope import snapshot_download, AutoModel, AutoTokenizer
import os</p><p>model_dir = snapshot_download('deepseek-ai/deepseek-moe-16b-chat', cache_dir='/root/autodl-tmp', revision='master')

python /root/autodl-tmp/download.py

from fastapi import FastAPI, Request
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
import uvicorn
import json
import datetime
import torch</p><h1>设备配置</h1><p>DEVICE = "cuda"
DEVICE_ID = "0"
CUDA_DEVICE = f"{DEVICE}:{DEVICE_ID}" if DEVICE_ID else DEVICE</p><h1>GPU内存清理函数</h1><p>def torch_gc():
if torch.cuda.is_available():
with torch.cuda.device(CUDA_DEVICE):
torch.cuda.empty_cache()
torch.cuda.ipc_collect()</p><h1>初始化FastAPI应用</h1><p>app = FastAPI()</p><p>@app.post("/")
async def create_item(request: Request):
global model, tokenizer
json_post_raw = await request.json()
json_post = json.dumps(json_post_raw)
json_post_list = json.loads(json_post)</p><pre class="brush:php;toolbar:false;"><pre class="brush:php;toolbar:false;">prompt = json_post_list.get('prompt')
max_length = json_post_list.get('max_length')

# 构建对话历史
messages = [
    {"role": "user", "content": prompt}
]

# 编码输入
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")

# 生成输出
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=max_length)
result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)

now = datetime.datetime.now()
time = now.strftime("%Y-%m-%d %H:%M:%S")

# 返回响应结构
answer = {
    "response": result,
    "status": 200,
    "time": time
}

# 打印日志
log = "[" + time + "] " + '", prompt:"' + prompt + '", response:"' + repr(result) + '"'
print(log)

torch_gc()
return answer

<code># 加载分词器
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

# 加载模型，启用bfloat16降低显存占用，自动分配GPU资源
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# 加载生成配置
model.generation_config = GenerationConfig.from_pretrained(model_path)
model.generation_config.pad_token_id = model.generation_config.eos_token_id

model.eval()  # 启用评估模式

# 启动API服务
uvicorn.run(app, host='0.0.0.0', port=6006, workers=1)</code></pre><h4>启动 API 服务</h4><p>在终端执行以下命令运行服务脚本：</p><pre><code class="javascript">cd /root/autodl-tmp

curl -X POST "<a href="https://www.php.cn/link/6190df6b9dfadcb413f0d0e3b768888a">https://www.php.cn/link/6190df6b9dfadcb413f0d0e3b768888a</a>" \
-H 'Content-Type: application/json' \
-d '{"prompt": "你好,你是谁？","max_length":100}'

import requests
import json</p><p>def get_completion(prompt, max_length):
headers = {'Content-Type': 'application/json'}
data = {"prompt": prompt, "max_length": max_length}
response = requests.post(url='<a href="https://www.php.cn/link/6190df6b9dfadcb413f0d0e3b768888a">https://www.php.cn/link/6190df6b9dfadcb413f0d0e3b768888a</a>', headers=headers, data=json.dumps(data))
return response.json()['response']</p><p>if <strong>name</strong> == '<strong>main</strong>':
print(get_completion("你好,你是谁？", 100))