Summary: 3 instances, 3 unique

Text	Count
# TODO: Use CUDA streams?	1
# TODO: Decide if we really want that for simple tensors?	1
# TODO: This isn't guaranteed to be exact if inference_batch_size	1