train/learner.py (3 lines):
	- line 617: # TODO: Decide if we really want that for simple tensors?
	- line 861: # TODO: Use CUDA streams?
	- line 866: # TODO: This isn't guaranteed to be exact if inference_batch_size