Upload app.py with huggingface_hub

app.py

@@ -5,7 +5,6 @@ from peft import PeftModel
 from PIL import Image
 import requests
 from io import BytesIO
-import gc
 
 model = None
 processor = None
@@ -13,23 +12,19 @@ processor = None
 def load_model():
     global model, processor
     if model is None:
-        print("Загружаю базовую модель Qwen2.5-VL-7B-Instruct...")
-        gc.collect()
-        torch.cuda.empty_cache() if torch.cuda.is_available() else None
+        print("Загружаю модель...")
         
         base_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
             "Qwen/Qwen2.5-VL-7B-Instruct",
-            torch_dtype=torch.bfloat16,
+            torch_dtype=torch.float16,
             device_map="auto",
-            trust_remote_code=True,
-            attn_implementation="flash_attention_2"
+            trust_remote_code=True
         )
         
-        print("Применяю LoRA адаптер...")
+        print("Применяю LoRA...")
         model = PeftModel.from_pretrained(
             base_model,
-            "Stepan222/oem-fake-classifier-qwen2vl",
-            torch_dtype=torch.bfloat16
+            "Stepan222/oem-fake-classifier-qwen2vl"
         )
         model.eval()
         
@@ -37,79 +32,52 @@ def load_model():
             "Qwen/Qwen2.5-VL-7B-Instruct",
             trust_remote_code=True
         )
-        print("✅ Модель загружена!")
+        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 загрузки модели: {e}"
+        return f"Ошибка модели: {e}"
     
     try:
-        if image_url.startswith("http"):
-            response = requests.get(image_url, timeout=10)
-            image = Image.open(BytesIO(response.content)).convert("RGB")
-        else:
-            return "Введите URL изображения"
-    except Exception as e:
-        return f"Ошибка загрузки изображения: {e}"
+        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 listing for auto parts. Classify as OEM (original manufacturer) or FAKE (aftermarket/counterfeit).
+    prompt = f"""Analyze this eBay auto part listing. Is it OEM (original) or FAKE (aftermarket)?
 
 {text}
 
-Look at the image and text carefully. Respond with:
-- OEM: [confidence]% - [brief reason]
-- FAKE: [confidence]% - [brief reason]"""
+Reply: OEM or FAKE with confidence % and reason."""
 
-    messages = [
-        {
-            "role": "user",
-            "content": [
-                {"type": "image", "image": image},
-                {"type": "text", "text": prompt}
-            ]
-        }
-    ]
+    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)
+    inputs = processor(text=[text_input], images=[image], return_tensors="pt", padding=True).to(model.device)
     
     with torch.no_grad():
-        outputs = model.generate(
-            **inputs,
-            max_new_tokens=100,
-            do_sample=False
-        )
-    
-    response = processor.batch_decode(outputs, skip_special_tokens=True)[0]
-    if "assistant" in response.lower():
-        response = response.split("assistant")[-1].strip()
+        out = model.generate(**inputs, max_new_tokens=80, do_sample=False)
     
-    return response
+    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="URL изображения", placeholder="https://i.ebayimg.com/..."),
-        gr.Textbox(label="Название товара", placeholder="BMW Genuine OEM Part..."),
-        gr.Textbox(label="Описание (опционально)", placeholder="", lines=2)
+        gr.Textbox(label="Image URL"),
+        gr.Textbox(label="Title"),
+        gr.Textbox(label="Description")
     ],
-    outputs=gr.Textbox(label="Результат", lines=3),
-    title="OEM/Fake Classifier (Qwen2.5-VL + LoRA)",
-    description="Классификатор автозапчастей. Space засыпает через 5 минут бездействия для экономии.",
+    outputs=gr.Textbox(label="Result"),
+    title="OEM/Fake Classifier",
     allow_flagging="never"
 )
 
-if __name__ == "__main__":
-    demo.launch()
+demo.launch()