Update app.py

app.py

@@ -40,7 +40,7 @@ model = loaded_model_name = tokenizer = image_processor = context_len = my_gener
 def load_model_async(model_path, model_name):
     global tokenizer, model, image_processor, context_len, loaded_model_name, my_generator
     print(f"Loading {model_name} model ... ")
-    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, None, model_name)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, None, model_name, load_4bit=True)
     if "llava" in model_name.lower():
         loaded_model_name = "LLaVA"
     else: