This is weird. TensorFlow sometimes messes up when you mix CPU and GPU packages (or release and nightly packages) in the same virtual environment. Could you try doing the following steps in a fresh virtualenv? It’s basically avoiding the first install of tensorflow since you want to install tensorflow-gpu. Also, you don’t need the deepspeech package for training.
I have identified the cause and am testing a fix. It’s a regression from commit dc78f8d1e65c4b387aa335df243898cd04704f98, which caused us to fallback on the default value for allow_soft_placement, which apparently is not True