Proposed solution doesn't seem to work anymore

[Paskin]

With Host OS 2.107.13+rev1, as well as 2.106.8 - rmmod of Nouveau module causes kernel exception in it (infamous nvkm_falcon_v1_wait_for_halt), leaving GPU in some intermediate state and crashing Nvidia original drivers loaded later.

[alanb128]

@Paskin are you referring to the Generic x86_64 (GPT) device type? Was 2.105.32+rev2 the last version that worked for you?

[Paskin]

are you referring to the Generic x86_64 (GPT) device type?
Yes

Was 2.105.32+rev2 the last version that worked
Did some research - no version with 5.x kernel seems to work. The only "hack" I found is to remove nouveau module from the installation flash image (unfortunately - do not remember why blacklisting wasn't helping, and do not have this machine anymore).

I wish Balena will produce an image with official support for nvidia-contatiner-toolkit etc one day (especially for Jetsons, where lack of GPU support invalidates the major point of the platform) - letting customers to run an unmodified containers from NVCR.

[closetelevator]

Does anyone from Balena have an update on this?

[alanb128]

Sometimes the Nouveau module can't be unloaded because Plymouth is using it. There's an example of how to work around this on our forums that may be helpful: https://forums.balena.io/t/blacklist-drivers-in-host-os/163437/25