I am running the model on a batch of 500 articles and it takes around 10 to 15 GB of GPU memory. Is there a way to optimize the GPU usage?