app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch
import re
import os

# 模型配置
MODEL_ID = "meta-llama/Llama-3.2-1B-Instruct"
ADAPTER_ID = "zhman/llama-SFT-GRPO"

# 全局变量存储模型和tokenizer
model = None
tokenizer = None

def load_model():
    """加载模型和tokenizer"""
    global model, tokenizer
    
    print("正在加载模型...")
    
    # 加载tokenizer
    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    
    # 加载基础模型
    base_model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        torch_dtype=torch.bfloat16,
        device_map="auto",
        trust_remote_code=True
    )
    
    # 加载 LoRA 适配器
    model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
    model.eval()
    
    print("模型加载完成!")
    return model, tokenizer

def extract_boxed_answer(text):
    """提取 \\boxed{} 格式的答案"""
    # 查找 \boxed{} 格式
    boxed_pattern = r'\\boxed\{([^}]+)\}'
    matches = re.findall(boxed_pattern, text)
    
    if matches:
        return matches[-1].strip()
    
    # 尝试其他格式
    patterns = [
        r'答案[:：]\s*([^\n]+)',
        r'Answer[:：]\s*([^\n]+)',
        r'= *([^\n]+)',
        r'因此[:：]\s*([^\n]+)',
        r'所以[:：]\s*([^\n]+)',
    ]
    
    for pattern in patterns:
        matches = re.findall(pattern, text)
        if matches:
            return matches[-1].strip()
    
    return None

def predict(question, max_new_tokens=1024, temperature=0.7):
    """
    模型推理函数
    
    Args:
        question: 数学问题
        max_new_tokens: 最大生成token数
        temperature: 温度参数
    
    Returns:
        (完整输出, 提取的答案)
    """
    global model, tokenizer
    
    # 首次调用时加载模型
    if model is None or tokenizer is None:
        load_model()
    
    # 构建prompt
    prompt = f"User: {question}\nPlease reason step by step.\nAssistant:"
    
    # Tokenize
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=1024)
    
    # 生成
    with torch.no_grad():
        outputs = model.generate(
            inputs.input_ids.to(model.device),
            max_new_tokens=max_new_tokens,
            temperature=min(temperature, 0.01) if temperature <= 0 else temperature,
            do_sample=temperature > 0,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id
        )
    
    # 解码
    full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
    response = full_output.replace(prompt, "").strip()
    
    # 提取答案
    answer = extract_boxed_answer(response)
    
    return response, answer if answer else "未能提取到答案"

# 创建 Gradio 界面
with gr.Blocks(title="数学问题求解 API") as demo:
    gr.Markdown("# 🧮 数学问题求解 API 后端")
    gr.Markdown("基于 Llama-3.2-1B + SFT + GRPO 微调模型")
    
    with gr.Row():
        with gr.Column():
            question_input = gr.Textbox(
                label="数学问题",
                placeholder="例如: 求解方程 x^2 + 5x + 6 = 0",
                lines=5
            )
            with gr.Row():
                max_tokens = gr.Slider(
                    minimum=128,
                    maximum=2048,
                    value=1024,
                    step=128,
                    label="最大生成长度"
                )
                temp = gr.Slider(
                    minimum=0.1,
                    maximum=1.5,
                    value=0.7,
                    step=0.1,
                    label="Temperature"
                )
            submit_btn = gr.Button("求解", variant="primary")
        
        with gr.Column():
            reasoning_output = gr.Textbox(
                label="推理过程",
                lines=15,
                max_lines=20
            )
            answer_output = gr.Textbox(
                label="提取的答案",
                lines=2
            )
    
    # 示例
    gr.Examples(
        examples=[
            ["Find the positive integer n such that 10^n cubic centimeters is the same as 1 cubic kilometer."],
            ["求解方程 3×5 等于多少？"],
        ],
        inputs=question_input
    )
    
    submit_btn.click(
        fn=predict,
        inputs=[question_input, max_tokens, temp],
        outputs=[reasoning_output, answer_output],
        api_name="predict"  # 重要: 启用 API 访问
    )
    
    gr.Markdown("---")
    gr.Markdown("### API 使用说明")
    gr.Markdown("""
    **API 端点**: `/api/predict`
    
    **POST 请求示例**:
    ```python
    import requests
    
    response = requests.post(
        "https://YOUR_SPACE_URL/api/predict",
        json={
            "data": [
                "你的数学问题",  # question
                1024,           # max_new_tokens
                0.7             # temperature
            ]
        }
    )
    result = response.json()
    reasoning = result["data"][0]  # 推理过程
    answer = result["data"][1]     # 提取的答案
    ```
    """)

# 启动应用
if __name__ == "__main__":
    demo.launch()