ubuntu-drivers-common
ubuntu-drivers-common copied to clipboard
tseliot
Require something like RUNTIMEPM_OVERRIDE to force install nvidia driver
Since there are many latest GPUs need to support, request ubuntu-drivers-common frequency release it not quite make sense.
Could we consider to have a backdoor to force install nvidia driver (after nvidia confirms the driver is support through business model) something like what RUNTIMEPM_OVERRIDE did to force on-demand mode.
In the other hand, if we force install nvidia driver for specifc GPU then we also need to take care about on-demand mode is support or not. If we could have a white list to install nvidia driver + switch nvidia mode then it will be better.
The main problem with this approach is that, installing a driver, is more complex than enabling RTD3. For example, which driver release shall be force installed? The current approach relies on matching the modaliases with the packages, so this would require rewriting the way we do the matching, which won't be as simple as the RUNTIMEPM_OVERRIDE thing.
I would like to socialize this idea.
It could benefit in this case https://bugs.launchpad.net/stella/+bug/1950553/comments/24
They are many GPU IDs are pre-release (nvidia don't want to put them in supported-gpu.json.
Usually, it needs someone to add it to nvidia-driver manually but it causes respin package effort.
If we can have a way to force install then it can reduce much efforts.
Says, the needed owner to put the ID to a file for u-d-c to reference.
If the concern is about the version of IDs, then how about:
2438:510
We can still search linux-module-nvidia-???-{kernel} nvidia-driver-??? there but have a new logic to force install if found.
If the idea is fine then I can prepare a MP for discussion.
What happens if, say, the 510 series is replaced by something newer in the archive?
do you mean 510 be transited to something like 530?
the idea from my mind is to check nvidia-driver-510 (or pre-signed) whether available if the package is there (even the transit package) then installing it (by some logic as ubuntu-drivers install nvidia:510).
It means if someone gets the confirmation that the GPU supported by a version (e.g. 510 in this case) from nVidia then the next version should backward compatible with previous version.
Also, the idea is for some GPU IDs not yet in supported-gpu.json.
When the GPU ID presents in supported-gpu,json, the author should remove it.
TBH, I don't think the normal user will use this method but developers (e.g. Canonical, nVidia, OEM/ODM)