try handle

handler.py

@@ -463,14 +463,20 @@ def initialize_model():
         print(f"Failed to initialize model: {e}")
         return False
 
-# Initialize model on import
-model_initialized = initialize_model()
+# Don't initialize model on import - do it lazily
+model_initialized = False
 
 # Main endpoint function for Hugging Face
 def query(payload):
     """Main endpoint function for Hugging Face inference API"""
+    global model_initialized
+    
+    # Lazy initialization - initialize model on first call
     if not model_initialized:
-        return {"error": "Model not initialized"}
+        print("Initializing model on first query...")
+        model_initialized = initialize_model()
+        if not model_initialized:
+            return {"error": "Model initialization failed"}
     
     try:
         # Extract parameters from payload
@@ -506,13 +512,18 @@ def health_check():
         "cuda_available": torch.cuda.is_available(),
         "llava_available": LLAVA_AVAILABLE,
         "transformers_available": TRANSFORMERS_AVAILABLE,
-        "cv2_available": CV2_AVAILABLE
+        "cv2_available": CV2_AVAILABLE,
+        "lazy_loading": True  # Model will be loaded on first query
     }
 
 def get_model_info():
     """Get model information"""
     if not model_initialized:
-        return {"error": "Model not initialized"}
+        return {
+            "error": "Model not initialized yet",
+            "lazy_loading": True,
+            "note": "Model will be loaded on first query"
+        }
     
     return {
         "model_path": args.model_path if args else "Unknown",