Many packages build on TensorFlow. For example, our work in Edward uses tensorflow>=1.0.0a0 as an install requirement.

However, this conflicts with tensorflow-gpu, which can no longer be installed because of the requirement specifically on tensorflow. What do you suggest is the best way to handle this?

One option suggested by @gokceneraslan (blei-lab/edward#428 (comment)) is to hack in the dependency according to whether the user has a GPU. Another option, which PrettyTensor and Keras employ, is to not even require TensorFlow. (Both options sound not good.)

Also see blei-lab/edward#428. also looping in GPflow devs (@jameshensman, @alexggmatthews) in case they have the same problem. (Note I'm raising this as an issue instead of asking on a mailing list, in case this is something that should be changed on TensorFlow's end and not our end.)