# vGPU variant: VM-only GPU workloads. Disables the pod-side driver # and device plugin (vfio-pci is not in play here), enables the # vGPU Manager DaemonSet. Image coordinates must be supplied by the # user via the Package CR — the proprietary .run cannot be # redistributed. gpu-operator: sandboxWorkloads: enabled: true # Without this override the upstream default 'container' applies on # GPU nodes that lack the per-node label # 'nvidia.com/gpu.workload.config'. In that case the operator picks # the container workload (driver / devicePlugin DaemonSets), which # are disabled below — so nothing GPU-related runs on the node and # vgpu-manager never starts. Setting the cluster-wide default to # 'vm-vgpu' makes the variant work out of the box. defaultWorkload: vm-vgpu driver: enabled: false devicePlugin: enabled: false vgpuManager: enabled: true repository: "" image: vgpu-manager version: "" # vGPU Device Manager walks /sys/class/mdev_bus/ to assign profiles. # On Ada Lovelace and newer (L4, L40, L40S, B100) with the vGPU 17/20 # driver branch this directory does not exist — the driver uses # per-VF NVIDIA sysfs (current_vgpu_type) instead, and the device # manager errors with «no parent devices found». Default-off keeps # modern hardware working out of the box; flip to true on # Pascal/Volta/Turing/Ampere (V100, T4, A100, A30, etc.) where the # mdev model still applies. vgpuDeviceManager: enabled: false