Please integrate ubuntu-drivers --gpgpu into Ubuntu Server

[dann frazier]

Could subiquity provide an option in the UI to install and execute ubuntu-drivers-common on the target? The use case I'm interested in is an "on-rails" installation of Nvidia drivers for servers being installed for deep learning workloads.

[Michael Hudson-Doyle (mwhudson)]

We should do this, but after 20.04 GA.

[dann frazier (dannf)]

I've found that ubiquity does this by calling ubuntu-drivers install (used to be autoinstall, but that changed). However, in server believe we'll want to add the --gpgpu option. This should cause u-d to prefer Nvidia drivers versions that are recommended for GPGPU purposes, as well as cause it to select the pre-built/signed driver versions vs. the DKMS ones.

[Dimitri John Ledkov (xnox)]

I don't think that implementing just subiquity is enough. Ubiquity was the last piece of ubuntu-drivers integration on the Desktop, as update-manager continiously was already checking for any new drivers that are available for the systems' hardware, and continiously providing notifications to install / switch between new alternatives / dismiss installing drivers, on continuous basis post-install / post-upgrade. Doing this during install, simply moved that step from post-install to pre-install.

Ubuntu Server currently does not have ubuntu-drivers integration with motd, unattanded upgrades, and apt output to notify users of the newly available driver implementations, and how to opt into them.

Command line tooling exists, but currently only used non-interactively by cloud-init on some compatible public clouds.

[Lee Trager (ltrager)]

MAAS contains a way to automatically install drivers. Every region has a file, /etc/maas/drivers.yaml, which specifies drivers which should be automatically installed. I don't have access to any test hardware but I *think* this should work. One thing I noticed is there isn't a meta package for the nVidia driver which points to the latest version for each Ubuntu release. We would need that before adding this to MAAS.

diff --git a/drivers-orig.yaml b/drivers.yaml
index 2d3724c..97a4eb2 100644
--- a/drivers-orig.yaml
+++ b/drivers.yaml
@@ -82,3 +82,10 @@ drivers:
   repository: http://downloads.linux.hpe.com/SDR/repo/ubuntu-hpdsa
   series:
     - trusty
+- blacklist: nouveau
+ comment: nVidia driver
+ modaliases:
+ - 'pci:v000010DEd*sv*sd*bc03sc02i00*'
+ - 'pci:v000010DEd*sv*sd*bc03sc00i00*'
+ - module: nvidia
+ - package: nvidia-headless-440

[Dan Watkins (oddbloke)]

cloud-init has support for ubuntu-drivers, it's not clear from this bug what work we would need to do so I've marked our task Incomplete.

[Jerzy Husakowski (jhusakowski)]

Unclear what needs to happen in MAAS, and if there's an issue with the current behavior.

[Michael Hudson-Doyle (mwhudson)]

The version of subiquity in 22.04.1 supports this now.

[James Falcon (falcojr)]

Tracked in Github Issues as https://github.com/canonical/cloud-init/issues/3649