GitLab

On ppc64, the installer asks confirmation to install without /boot.
It's not a problem, so add the answer to the preseed configuration.

This should be used for any software that builds kernel modules, so that
it pulls the correct version of linux-headers.

Fix nvidia for ppc64

See merge request !72

Fixes: 26bab2b0 ("Correctly set the kernel to use when creating Debian images")

Correctly set the kernel to use when creating Debian images

See merge request !70

It can happen that after the initial upgrade, a kernel upgrade occures.
In such case, the links /{vmlinuz,/initrd.img} point to the "old" kernel
version. Also, the Nvidia driver is compiled for the currently running
kernel.
Now, the last kernel while be used to correct both cases.

dcgm-exporter links against datacenter-gpu-manager 1.X, so it does not
find the .so file of datacenter-gpu-manager 2.X.  Force version 1.X.

Both cuda and prometheus have services that need to start only on machines
with GPU.  They rely on /dev/nvidia0 being present on such machines to
detect the presence of GPU.

However, /dev/nvidia0 is not created when the nvidia driver is loaded, but
only when `nvidia-smi` is first run.

The new service calls `nvidia-smi` once during boot to create /dev/nvidia0
and thus to allow the other services to start.  On machines without GPU,
`nvidia-smi` will fail, so we ignore its exit code.

This was apparently never done, so the service is installed since 2018 but
never actually run...

We only start the service if /dev/nvidia0 exists (i.e nvidia driver is
loaded), because otherwise the service would fail.