Update app.py

app.py

@@ -23,7 +23,7 @@ device = "cuda"
 device_map = "auto"
 
 # Load model, tokenizer, and image processor
-tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map)
+tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, device_map=device_map, attn_implementation="sdpa")
 model.eval()
 
 def respond(