System update wants to remove cuda-10 and related, seemingly w/ no replacement

I’m a university researcher, and I spent a good deal of time compiling Tensorflow GPU to work w/ CUDA 10 (doesn’t install automatically) and adding a number of packages by hand to get a working neural network system on my computer. I’m running Ubuntu 18.04.

Today, when I went to do an everyday update with apt, I was told that the system wanted to remove:

cuda-10-0 cuda-demo-suite-10-0 cuda-drivers cuda-runtime-10-0 libnvidia-compute-410 libnvidia-decode-410 libnvidia-encode-410
libnvidia-fbc1-410 libnvidia-gl-410 libnvidia-ifr1-410 nvidia-compute-utils-410 nvidia-dkms-410 nvidia-driver-410
nvidia-kernel-common-410 nvidia-utils-410

A bunch more CUDA-related packages would no longer be utilized.

The only new package my system wants to install is: libnvidia-compute-418

And it wants to upgrade the following:

dh-translations gir1.2-dbusmenu-glib-0.4 gir1.2-gtk-3.0 gtk-update-icon-cache libdbusmenu-glib4 libdbusmenu-gtk3-4 libdbusmenu-gtk4
libgail-3-0 libgtk-3-0 libgtk-3-bin libgtk-3-common libx11-6 libx11-data libx11-dev libx11-doc libx11-xcb-dev libx11-xcb1 libxnvctrl0
nvidia-modprobe nvidia-settings unattended-upgrades

I can’t find information on libnvidia-compute-418, but am guessing that it’s not going to substitute for cuda-10 by itself. It will likely just substitute for libnvidia-compute-410, which is one of the packages to be removed.

I see couple possibilities: Maybe once I install the above packages, the next update will call in some update to cuda 10 that will fit into my tensorflow setup (or break it). OR maybe I should try to hold back my CUDA 10 etc. to avoid changes that may break my neural network software.

Any info., thoughts would be welcome.


I suppose one solution to this issue is to move to the recommended Nvidia Docker virtual environment, so updates to my host system won’t affect my Tensorflow setup. I wonder, though, how well the Docker environment works–is it fast, easy to work with? I have a big Anaconda setup on my host system, which I’m guessing I’d now have to replicate in Docker.

Having my computer in a restart state, despite not applying changes, was causing problems, so I went ahead and accepted the changes.

I discovered that there’s a new version of CUDA, CUDA-10-1. For whatever reason, my apt update didn’t try to install that, so I chose it by hand. That replaced, I believe, all of the software that the update wanted to erase, with newer versions.

Whether the new version of CUDA works w/ my hand-compiled version of tensorflow-gpu, I don’t yet know. I’m hoping it does or that there is now a tensorflow-gpu I can easily install that works with CUDA-10-1. If not, then I’m back to considering Docker.

If cuda-10.1 doesn’t work out for you (I doubt that it will since tensorflow 1.13.1 just released 6 days ago is the only version that works with cuda 10.0, which I compiled from source), you can remove cuda from synaptic, then install cuda-10-0, then apt will not ask you to upgrade. If you don’t have synaptic, install it. It is a convenient package managing GUI so you can see all cuda versions available at a glance.

Thanks Bernard! My home system didn’t go through this hiccup evidently because I had installed a cuda-10-0 repository directly into synaptic.

The info that tensorflow just caught up w/ cuda-10 is very helpful as is the tip that I can just reinstall cuda-10 w/o worrying about repositories.

Much appreciated