I have been using a trained model to annotate a larger number of abstracts but I see that after some time, there is a large number of threads that are not being closed. It happens when I run it on a CPU or GPU. Are you aware of any solution for this problem?