Changes for version 0.004 - 2026-10-01
- Not run on GPU hardware yet; checked against the commands sent to the host.
- CDI
- generate_cdi_specs checks with `nvidia-ctk cdi list` that nvidia.com/gpu=all and management.nvidia.com/gpu=all resolve, and writes only what is missing under /etc/cdi: nvidia.yaml, and management.nvidia.yaml via --mode=management --vendor=management.nvidia.com --class=gpu. Fixes hosts without nvidia-cdi-refresh (DGX OS / GB10, toolkit 1.19), where /run/cdi is empty after a reboot and the GPU Operator validator failed on "unresolvable CDI devices management.nvidia.com/gpu=all". Also in the "toolkit already present" case. Specs that resolve are left alone (also nvidia-cdi-refresh's under /run/cdi); next to nvidia-cdi-refresh no static nvidia.yaml is written. A CLI without the management mode, or without `cdi list`, is a warning, not a die. An existing /etc/cdi/nvidia.yaml is no longer rewritten on every run (k76).
Modules
GPU detection and driver management for Rex
GPU hardware detection via PCI class codes
GPU detection from sysfs, without lspci or pciutils
NVIDIA GPU driver and container toolkit management
What NVIDIA driver a GPU generation needs (experimental)
Base class of the per-distro NVIDIA driver setups (experimental)
apt/dpkg packaging layer of the NVIDIA driver setups (experimental)
NVIDIA driver setup for Debian (experimental)
NVIDIA driver setup for RHEL, Rocky, AlmaLinux and CentOS Stream (experimental)
rpm packaging layer of the NVIDIA driver setups (experimental)
NVIDIA driver setup for openSUSE Leap (experimental)
NVIDIA driver setup for Ubuntu (experimental)
NVIDIA driver setup for Ubuntu, package named by ubuntu-drivers (experimental)
Which PCI subsystem IDs mark an NVIDIA vGPU guest device