Installing nvidia drivers and cuda toolkit versions that are compatible with tensorflow is proving to be as enjoyable as ingrown toenail surgery.
I specifically need cuda toolkit 10.0 because tensorflow 1.14 only supports this version. If I could use 10.1 or 10.2, I would have solved this weeks ago.
I have followed numerous posts [1][2][3][4][5] to try and uninstall all existing cuda and nvidia software and start again from scratch, but nothing works; sometimes I get Unmet Dependancies errors and when I try to install the dependancies, I get new errors until I am 5 dependancies down a rabbit hole.
The latest I have tried is to manually download cuda 10.0 from here and run the .sh file (runfile (local)). I select Y for all the prompts except for installing the samples. This seems to work fine; I see cuda 10.0 is installed under /usr/local/. Then I install the nivida drivers using
$ sudo apt-get install --no-install-recommends nvidia-driver-430
here is where the problems start. I reboot the computer, run nividia-smi and find that while the driver is installed correctly (version 430.50), the toolkit is 10.1!!! It seems that installing the driver is overwriting the cuda toolkit version? cuda-10.0 is still located under /usr/local/ and if I run /usr/local/cuda/bin/nvcc --version or cat /usr/local/cuda/version.txt I see that both give version 10.0. I don't understand how the nvidia driver is using (or at least displaying) cuda version 10.1.
Is there some way I can install cuda toolkit version 10.0 and some nvidia driver (>=410.x) so that tensorflow will work with my GPU?
