ubuntu-drivers --gpgpu is not installing -server packages

[Jeff Lane ]

Discovered while testing the ubuntu-drivers-common package in -proposed for bug LP:1898601

While waiting on bionic to deploy for verification, something was bugging me about Focal verification from the other day so I re-deployed a DL380 with 4x T4 GPGPUs with Focal and added u-d-c from proposed. What I missed when verifying the other day was that it actually isn't installing the -server package on my server in gpgpu mode. After chatting a bit with Alberto, --gpgpu is supposed to install the nvidia-driver-VER-server package. Instead, the version in -proposed is actually installing nvidia-driver-VER which is incorrect.

Attached is a tarball from a fresh focal install that includes the output when doing 'ubuntu-drivers --gpgpu install' and 'ubuntu-drivers --gpgpu debug' after installing the driver.

[Jeff Lane  (bladernr)]

This is still broken as of 1:0.8.6.3~0.18.04.2

$ ubuntu-drivers --gpgpu devices
This is gpgpu mode
== /sys/devices/pci0000:c9/0000:c9:02.0/0000:ca:00.0/0000:cb:01.0/0000:cd:00.0 ==
modalias : pci:v000010DEd000020B5sv000010DEsd00001533bc03sc02i00
vendor : NVIDIA Corporation
driver : nvidia-driver-470-server - distro non-free
driver : nvidia-driver-470 - distro non-free recommended
driver : xserver-xorg-video-nouveau - distro free builtin

This system has 8x A100 compute GPUs, is running Ubuntu Server, and I specified --gpgpu because I want the server compute drivers. But U/D is still insisting on installing the desktop graphics driver instead

[Jeff Lane  (bladernr)]

$ apt-cache policy ubuntu-drivers-common
ubuntu-drivers-common:
  Installed: 1:0.8.6.3~0.18.04.2
  Candidate: 1:0.8.6.3~0.18.04.2
  Version table:
 *** 1:0.8.6.3~0.18.04.2 500
        500 http://us.archive.ubuntu.com/ubuntu bionic-updates/main amd64 Packages
        100 /var/lib/dpkg/status
     1:0.5.2 500
        500 http://us.archive.ubuntu.com/ubuntu bionic/main amd64 Packages

[Jeff Lane  (bladernr)]

https://pastebin.ubuntu.com/p/nzxSG3DMVF/

This remains broken in ubuntu-drivers. There is absolutely no difference between what is installed when doing

ubuntu-drivers install

and

ubuntu-drivers install --gpgpu

but significant difference when using

ubuntu-drivers install nvidia:470-server

the first two install exactly the same package set and the latter installs a much smaller subset.

If you actually expect that the first two install the exact same package set, why does --gpgpu exist in the first place? If the expectation is that --gpgpu means VER-server, then this remains broken since last year. And if the expectation exists in the middle, then perhaps we need a --gpgpu-server to install the packages from the last example?

I guess the first question for this bug is "What EXACTLY is expected when installing using --gpgpu" and then "how do we resolve this bug"