Note about PyTorch CPU-only builds
Many Python-based LT tools make use of PyTorch, and the default torch
packages on PyPi are very large (>700MB) as they include all the code for GPU-based training and inference. On ELG, at least at present, we only do CPU-based inference, so service providers can save 5-600MB of image size by using alternative PyTorch builds that do not include all the GPU baggage.