I have been working on a model whose training loop uses a tf.function wrapper (I get OOM errors when running eagerly), and training seems to be running fine