app.py

import gradio as gr
import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from peft import PeftModel
from PIL import Image
import requests
from io import BytesIO

model = None
processor = None

def load_model():
    global model, processor
    if model is None:
        print("Загружаю модель...")
        
        base_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
            "Qwen/Qwen2.5-VL-7B-Instruct",
            torch_dtype=torch.float16,
            device_map="auto",
            trust_remote_code=True
        )
        
        print("Применяю LoRA...")
        model = PeftModel.from_pretrained(
            base_model,
            "Stepan222/oem-fake-classifier-qwen2vl"
        )
        model.eval()
        
        processor = AutoProcessor.from_pretrained(
            "Qwen/Qwen2.5-VL-7B-Instruct",
            trust_remote_code=True
        )
        print("Модель готова!")
    return model, processor

def classify(image_url: str, title: str, description: str = ""):
    try:
        model, processor = load_model()
    except Exception as e:
        return f"Ошибка модели: {e}"
    
    try:
        response = requests.get(image_url, timeout=10)
        image = Image.open(BytesIO(response.content)).convert("RGB")
    except:
        return "Не удалось загрузить изображение"
    
    text = f"Title: {title}"
    if description:
        text += f"\nDescription: {description}"
    
    prompt = f"""Analyze this eBay auto part listing. Is it OEM (original) or FAKE (aftermarket)?

{text}

Reply: OEM or FAKE with confidence % and reason."""

    messages = [{"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": prompt}]}]
    
    text_input = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = processor(text=[text_input], images=[image], return_tensors="pt", padding=True).to(model.device)
    
    with torch.no_grad():
        out = model.generate(**inputs, max_new_tokens=80, do_sample=False)
    
    resp = processor.batch_decode(out, skip_special_tokens=True)[0]
    return resp.split("assistant")[-1].strip() if "assistant" in resp.lower() else resp

demo = gr.Interface(
    fn=classify,
    inputs=[
        gr.Textbox(label="Image URL"),
        gr.Textbox(label="Title"),
        gr.Textbox(label="Description")
    ],
    outputs=gr.Textbox(label="Result"),
    title="OEM/Fake Classifier"
)

demo.launch()