Yeah, ML libraries are a huge offender. The Torch 1.8.1 manylinux wheel is 1.6G unpacked, with 1.1G going to CUDA support.