Can you expand on point 4? Conda-forge releases perfectly fine tensorflow-gpu builds, with the caveat that they don’t ship stubs or the actual NVIDIA driver with them so it’s not truly standalone, but the same can be said of pytorch or really any GPU-enabled package.
A particular build of Tensorflow X requires version A of the CUDA library, version B of the CDNN library, etc.
It is a common situation if you work on a data science team or want to play with models you find on Github that some of them require X1, A1, B1 and some others require X2, A2, B2.
The CUDA and cuDNN libraries are ordinary userspace libraries so if you package them for anaconda you can install them into a virtualenv and have different versions of the libraries sitting side by side and never get an error because the library versions don't match -- and I've done that on both Windows and Linux.
Anaconda can't ship conda packages like the ones I describe because NVIDIA insists that you download the libraries from their website, register to get senseless spam, screw around with installers, etc.
