Rex-GPU-0.002

Changes for version 0.002 - 2026-09-24

None of the changes below has run on GPU hardware yet; they are checked against the commands they send to the host.
Detection
- A GPU counts as compute by its PCI device ID generation, not its name: every Maxwell or newer NVIDIA GPU (GeForce included) now gets a driver. Kepler and older are compute => 0 with a warning.
- A virtual display (virtio, QXL, VMware, VirtualBox) no longer hides a real GPU on passthrough hosts and cloud GPU VMs.
- New keys: nvswitch (HGX NVSwitch bridges) and per GPU vgpu, vgpu_type, subsystem_vendor_id, subsystem_id (NVIDIA vGPU guests).
- pciutils is installed only when lspci is missing, also on Rocky, Alma and CentOS Stream; detection dies if lspci is still missing.
Driver
- gpu_setup installs one driver that fits all compute GPUs (new option gpus => [...]). It dies before any driver package is installed when they cannot share one, for a Kepler GPU, or for a vGPU guest without a working driver.
- The driver fits the GPU generation: Blackwell (B200/B300, RTX 50xx, RTX PRO, GB10) gets the open kernel module, on Debian 12/13 from NVIDIA's CUDA repository; Maxwell/Pascal/Volta stay on the proprietary 580 branch. GB10 (DGX Spark) needs 580 or newer.
- A host with a working driver (nvidia-smi lists a GPU, libcuda is present) is left alone: no reinstall, no reboot.
- HGX hosts with NVSwitch get NVIDIA Fabric Manager; HGX B200/B300 also get the NVLink fabric (nvlsm, ib_umad). An existing driver gets a missing Fabric Manager from the host's own package sources.
- Ubuntu: the newest nvidia-driver-NNN-server package is picked after apt-get update; no silent fallback to 570.
- Debian: contrib/non-free are enabled on every Debian archive entry, deb822 .sources included. An own mirror counts when it is signed-by the Debian archive keyring.
- RHEL family: Rocky, Alma and CentOS Stream work under their lsb_release names; RHEL gets EPEL and CodeReady Builder; aarch64 uses NVIDIA's sbsa repository; a CUDA repository that cannot be added dies early.
- reboot => 1 waits until the host answers a command again.
- Own driver setup: setup => 'My::Setup' or set gpu_nvidia_setup, plus requirement => { kernel_module, min_branch, max_branch }. Setup classes and Rex::GPU::NVIDIA::Requirement are experimental API; examples in eg/custom-setup/ and eg/ubuntu-drivers/.
Container toolkit, CDI, containerd
- install_container_toolkit skips an installed toolkit, survives the dpkg lock after a reboot, and dies on a failed key or repository download instead of installing from a broken source.
- generate_cdi_specs leaves CDI to nvidia-cdi-refresh when the toolkit ships it, so nvidia.com/gpu is no longer defined twice.
- configure_containerd adds the nvidia runtime to RKE2/K3s's own config instead of replacing it (SystemdCgroup, sandbox image and registry config were lost). A template written by 0.001 is removed with a warning; restart rke2/k3s to regenerate the config.
- An unknown containerd_config dies before anything runs on the host.
openSUSE (unverified, unsupported)
- Leap 15.x gets the Leap 15 repository and G06 driver again (it got Leap 16's). zypper waits for the zypp lock, repository failures die, and every install is verified.
Examples
- eg/ keeps SSH host-key verification on and scans the new host into known_hosts first.