app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import json

# 全局变量，避免重复加载
model = None
tokenizer = None

def load_model():
    """加载模型和分词器"""
    global model, tokenizer
    model_name = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = AutoModelForCausalLM.from_pretrained(
            model_name,
            torch_dtype=torch.float16,  # 使用半精度减少内存占用
            device_map="auto",         # 自动分配至CPU
            low_cpu_mem_usage=True     # 优化CPU内存使用
        )
        print("模型加载成功！")
    except Exception as e:
        print(f"模型加载失败: {e}")

def openai_compatible_api(message, history):
    """处理OpenAI格式的请求"""
    if model is None:
        load_model()

    # 将历史记录和当前消息组合成对话格式
    # 这里需要根据DeepSeek模型要求的模板格式化工序
    # 以下是一个简化示例，您需要根据模型的具体要求调整
    prompt = f"\n\nHuman: {message}\n\nAssistant:"

    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=1024)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.7,
            top_p=0.95,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id
        )

    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    # 提取模型生成的部分
    generated_text = response.split("Assistant:")[-1].strip()

    # 返回OpenAI兼容格式
    return {
        "choices": [{
            "message": {
                "role": "assistant",
                "content": generated_text
            }
        }]
    }

# 在Gradio界面启动前加载模型（可选，或等待第一个请求时加载）
load_model()

# 创建Gradio界面，但隐藏默认的Web UI，专注于API
demo = gr.ChatInterface(
    fn=openai_compatible_api,
    title="DeepSeek API Service",
    description="OpenAI-compatible API for DeepSeek-R1"
)

# 禁用自动创建公共链接，仅以API模式运行
if __name__ == "__main__":
    demo.launch(show_api=True, server_name="0.0.0.0", server_port=7860)