1

I've been trying to install tensorflow-gpu, and as part of that process I need to install the cuda toolkit. Nvidia's instructions for doing so direct you to download a .deb file and install it, which I dutifully did. Next, they tell you to run sudo apt-get install cuda

I did so and the install went without a hitch, but the cuda package upgraded my drivers from nvidia-370 to nvidia-375. Now I'm dealing with a really nasty sleep bug which I assume is the fault of the new drivers. Therefore, I'd like to downgrade my drivers.

However, when I run sudo apt-get install nvidia-370, Ubuntu tells me that it will remove not only nvidia-375, but cuda and all its associated packages as well. And installing the cuda package automatically upgrades my graphics drivers.

Has anyone managed to get tensorflow/the cuda library installed on anything but the 375 drivers?

pipsqueaker
  • 340
  • have you got the graphics-drivers ppa installed? I've pinned mine to 367.57, and installed cuda 8.0.44. Worked on 16.04 and now 16.10. Do you need the 370 series for a Pascal card or something? – Andrew Keech Feb 17 '17 at 09:56
  • @AndrewKeech I do have the ppa installed. How did you pin your drivers, and how did you install Cuda? I was following nvidia's official instructions @ http://docs.nvidia.com/cuda/cuda-installation-guide-linux/#axzz4VZnqTJ2A – pipsqueaker Feb 17 '17 at 15:05
  • it's "pinned" in the sense that I haven't changed it in the system settings from 367.57. I used the .deb download from nvidia for cuda 8.0.44 (1.9 Gbs), and the cudnn 5.1.5 (39 Mbs), which is another sign-up and download iirc. Tensorflow was installed in a virtualenv, and didn't need any more convincing than that to work with it – Andrew Keech Feb 19 '17 at 10:02

0 Answers0