app.py

import os
import gradio as gr
import fitz
import docx
from huggingface_hub import InferenceClient

# 1. 引擎配置：Qwen-2.5-72B-Instruct (全能稳定版)
model_id = "Qwen/Qwen2.5-72B-Instruct"
hf_token = os.getenv("HF_TOKEN")

client = InferenceClient(
    model=model_id, 
    token=hf_token,
    # 增加超时时间设置，防止翻译长文时断连
    headers={"Authorization": f"Bearer {hf_token}"} if hf_token else None
)

def extract_text(file_path):
    if not file_path: return ""
    try:
        ext = str(file_path).lower()
        if ext.endswith(".pdf"):
            doc = fitz.open(file_path)
            return "".join([page.get_text() for page in doc])[:15000]
        elif ext.endswith(".docx"):
            doc = docx.Document(file_path)
            return "\n".join([p.text for p in doc.paragraphs])[:15000]
        elif ext.endswith(".txt"):
            with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
                return f.read()[:15000]
        return ""
    except:
        return ""

# 注意：这里移除了 @spaces.GPU，因为 API 模式下它反而会因超时导致报错
def chat_fn(message, history):
    clean_messages = []
    clean_messages.append({"role": "system", "content": "You are a professional patent and legal translation expert. Provide accurate translations."})
    
    for turn in history:
        clean_messages.append({"role": turn["role"], "content": str(turn["content"])})
    
    prompt_text = message.get("text", "")
    files = message.get("files", [])
    
    if files:
        f = files[0]
        file_path = f.get("path") if isinstance(f, dict) else f
        context = extract_text(file_path)
        if context:
            prompt_text = f"【参考文件】:\n{context}\n\n【指令】: {prompt_text}"
        
    clean_messages.append({"role": "user", "content": prompt_text})

    response = ""
    try:
        # 增加流式输出
        stream = client.chat_completion(
            clean_messages, 
            max_tokens=4096, 
            stream=True, 
            temperature=0.4
        )
        
        for msg in stream:
            if hasattr(msg, 'choices') and len(msg.choices) > 0:
                delta_content = msg.choices[0].delta.content
                if delta_content:
                    response += str(delta_content)
                    yield response
        
        if response:
            yield response + "\n\n---\n> **OpenIPOS MegaNode | Qwen-2.5 专业算力出口**"
            
    except Exception as e:
        yield f"⚠️ 算力调度提示: {str(e)}。由于翻译长文档耗时较长，请分段进行或刷新重试。"

# 3. 统一品牌装修
with gr.Blocks(fill_height=True) as demo:
    gr.HTML("""
        <div style="display: flex; align-items: center; gap: 20px; padding: 10px;">
            <img src="https://s3.bmp.ovh/2026/03/30/rIAGKg0O.png" style="height: 60px; width: auto;">
            <div>
                <h1 style="margin: 0; font-size: 24px;">OpenIPOS MegaNode</h1>
                <p style="margin: 0; color: #666;">全球顶尖中文全能算力出口 (Qwen-2.5-72B)</p>
            </div>
        </div>
    """)
    gr.ChatInterface(chat_fn, multimodal=True)
    gr.Markdown("© 2026 OpenIPOS Global Limited. 算力驱动：Novita API Cluster")

demo.launch(theme=gr.themes.Soft())