app.py

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

# --- 1. 配置与模型加载 ---
MODEL_ID = os.getenv("MODEL_ID", "badanwang/teacher_basic_qwen3-0.6b")
print(f"正在加载模型: {MODEL_ID}")

# 尝试加载模型，如果失败则在界面上显示错误
try:
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
    # 使用 device_map="auto" 让 accelerate 库自动处理设备分配
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        torch_dtype="auto",
        device_map="auto",
        trust_remote_code=True
    )
    print("模型和分词器加载成功！")
    
    # 定义核心推理函数
    def predict(prompt: str, history: list[list[str]]):
        """
        接收输入和历史，返回更新后的历史。
        Gradio 会自动为此函数创建 API 端点。
        """
        print(f"收到请求: prompt='{prompt}'")
        
        messages = []
        for user_message, bot_message in history:
            messages.append({"role": "user", "content": user_message})
            messages.append({"role": "assistant", "content": bot_message})
        messages.append({"role": "user", "content": prompt})

        input_ids = tokenizer.apply_chat_template(
            messages,
            add_generation_prompt=True,
            tokenize=True,
            return_tensors="pt"
        ).to(model.device)

        outputs = model.generate(input_ids, max_new_tokens=1024)
        response_text = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
        
        print(f"生成回复: {response_text}")
        
        history.append([prompt, response_text])
        return history

except Exception as e:
    print(f"加载模型时发生致命错误: {e}")
    # 如果模型加载失败，则定义一个报错函数
    def predict(*args, **kwargs):
        raise gr.Error(f"模型加载失败，请检查Space后台日志以确认是否为内存不足。错误详情: {e}")

# --- 2. 创建并启动Gradio应用 ---
with gr.Blocks(theme=gr.themes.Default()) as demo:
    gr.Markdown(f"## 简易模型聊天 ({MODEL_ID})")
    chatbot = gr.Chatbot(label="对话窗口", height=600)
    msg = gr.Textbox(label="输入你的问题")
    clear = gr.Button("清除对话")

    msg.submit(predict, [msg, chatbot], chatbot)
    clear.click(lambda: [], None, chatbot)

# .queue() 允许处理排队请求
# api_open=True 是关键，它会自动创建 /run/predict API 端点
print("准备启动Gradio应用...")
demo.queue().launch(api_open=True)