GPU device passthrough is an experimental feature available in commercial Deckhouse Platform (DP) editions.
DP attaches physical GPU devices to virtual machines using DRA (Dynamic Resource Allocation). A project owner requests a device by a reference to a GPUClass in the .spec.gpus block of their machine, and you prepare the cluster for this.
To make passthrough work, provide the following:
- Kubernetes 1.34 or later with the DRA feature gates required by your cluster configuration.
- The
GPUfeature gate in the module settings. - The enabled
gpumodule in the DRA mode, which is set by thedra.enabledparameter. - A
GPUClassresource that selects devices of the model you need. Thegpumodule creates a DeviceClass resource with the same name from it, and the device is allocated to a machine through that class.
To enable the feature gate, add it to the module settings:
apiVersion: deckhouse.io/v1alpha1
kind: ModuleConfig
metadata:
name: virtualization
spec:
settings:
featureGates:
- GPU
After that, tell the project owners the names of the available GPUClass resources. A single machine takes no more than 16 devices, and a change to the .spec.gpus block applies only after the machine restarts.