I received my DGX Spark. When initiating the Spark, the software update stalled at 80% for 6 hours, after which I terminated the process.
Please advise.
In which state is your spark right now? There were some problem with Ubuntu servers on the Spark the last 3 days.
I had some pretty terrible update experiences over the last few days. After a managed dashboard update, the NVIDIA driver package (nvidia-dkms-580-open) failed to build the DKMS kernel module for the new running kernel (6.17.0-1014-nvidia), leaving the driver in a half-configured state. Fix was to manually run sudo dkms install nvidia/580.142 -k 6.17.0-1014-nvidia then reboot. Second time in two days.
@ibrunton_smith DKMS is not part of DGX OS. The nvidia driver is packaged as a .deb. If you are using nvidia-dkms-580-open try switching back to the DGX OS default package, and disable dkms service if not used for any other modules.
@elsaco can you point me to a web resource for the revert? Not sure how I’ve managed to get myself off the standard path as I’ve only ever used the auto update through dashboard?!
Run dkms remove nvidia/580.142 -all to uninstall the driver, then reboot in text mode and install linux-modules-nvidia-580-open-6.17.0-1014-nvidia package. This is the package that has nvidia.ko and the related modules. It will pull all dependencies if not present already on your Spark.
If you don’t need dkms.service for anything else disable it, or uninstall the dkms package.
After the Nvidia driver is installed and working run systemctl set-default graphical.target and reboot, if you use the desktop.
The nvidia-driver-580-open is the metapackage for the 580 drivers.
Amazing! Thanks so much for taking the time to reply.
There was a temporary degradation in Canonical’s servers which may have interfered with some units setup.
I’ve had the NVIDIA driver break twice after managed dashboard updates on my DGX Spark. Both times the issue was the same: nvidia-dkms-580-open failed to build the kernel module for the new running kernel (6.17.0-1014-nvidia), leaving the driver in a half-configured (iU) state. I fixed it manually with sudo dkms install nvidia/580.142 -k 6.17.0-1014-nvidia and a reboot.
Following advice here, I tried switching to the prebuilt module (linux-modules-nvidia-580-open-6.17.0-1014-nvidia), but it won’t install because it expects nvidia-kernel-common-580 <= 580.126.09, while I have 580.142 from the canonical-nvidia/nvidia-desktop-edge PPA.
Two questions:
- Should the nvidia-desktop-edge PPA be enabled on a DGX Spark? It seems to be pulling in newer driver packages than the prebuilt kernel modules expect.
- Could the recent Canonical connectivity-check outage (21-23 March) have contributed to the failed updates?
Currently working via DKMS but would like to get onto the correct DGX OS driver path. Direct ref to @elsaco and @aniculescu for assistance here.
Just following up on this in case it is of use to others. I ended up getting into quite a mess when trying to fix an initial failed dashboard update which I suspect was caused by the issues noted above that broke nvidia-smi. Initial troubleshooting with AI (claude) suggested a standard Ubuntu solution (run `dkms autoinstall`) which got me out of sync with the approved packages.
After a fair bit of work this morning I think that I am now back on driver 580.126.09 with prebuilt kernel modules, no DKMS — matching the official DGX OS 7.4.0 release notes. Thanks to @elsaco and @aniculescu for the guidance, and a useful reminder not to blindly follow the first AI solution…