This is hell for a lot of ML containers, that have gigabytes of CUDA and PyTorch. Before at least you could keep your code contained to a layer. But if I understand this correctly every code revision duplicates gigabytes of the same damn bloated crap.