Recently stumbled on this problem. It seems like it basically makes impossible usage of compute with custom docker images for any pytorch-based real life computer vision ML experiments. Which is unfortunate. +1 for requesting followup and possible alternative solutions! Thank you!