So the method should be correct, but it seems that there are cases where an error occurs during merging if the model is quantized.