Rex-GPU-0.003

Changes for version 0.003 - 2026-10-01

None of the changes below has run on GPU hardware yet; they are checked against the commands they send to the host.
Driver
- Grace Hopper GH200 (10de:2342, 10de:2348) gets the open kernel module only, as Blackwell: Ubuntu -server-open, Debian 12/13 NVIDIA's CUDA repository, RHEL cuda-open-dkms, openSUSE its open driver.
- Ubuntu: only the running kernel's linux-headers-$(uname -r) are installed, no longer linux-headers-generic, which brought headers for another kernel on HWE and vendor kernels (DGX Spark). A kernel installed later needs its own headers for DKMS.
- Ubuntu: new opt-in setup Rex::GPU::NVIDIA::Setup::UbuntuDrivers (experimental): `ubuntu-drivers list --gpgpu` names the driver package, installed by apt-get and verified with dpkg; nothing named dies before any driver install. The default stays apt-cache search.
Detection
- An AMD GPU's name is read from lspci as for NVIDIA, e.g. "Navi 31 [Radeon RX 7900 XTX]", instead of "Unknown AMD GPU" for every card. Detect-only as before: AMD stays compute => 0.
- New opt-in Rex::GPU::Detect::Sysfs (experimental): detection from /sys/bus/pci/devices, without lspci and without installing pciutils; same result shape including vgpu and nvswitch. No product names, so a GPU the generation table does not cover is compute => undef, not 0. Dies where sysfs cannot be read. Chosen with detect => or set gpu_detect_class; lspci stays the default.
API
- Every function of Rex::GPU, Rex::GPU::Detect and Rex::GPU::NVIDIA also works as a class method, and a subclass of Rex::GPU::NVIDIA or Rex::GPU::Detect can replace any step (experimental): chosen with gpu_setup(nvidia => ..., detect => ...) or set gpu_nvidia_class / set gpu_detect_class; one that cannot be loaded dies before the host is touched. Function calls, their arguments and the defaults are unchanged.
Container toolkit
- install_container_toolkit(binaries_suffice => 1) counts the toolkit as present when nvidia-container-runtime and nvidia-ctk are on the PATH and nvidia-ctk runs, without asking the package manager (vendor images such as DGX Spark); gpu_setup and the default are unchanged