The module lifecycle stage: General Availability
The module has requirements for installation

Below are NodeGroup snippets for typical GPU modes. GPU-node management requires the NVIDIA driver and NVIDIA Container Toolkit installed on the nodes (manually or via NodeGroupConfiguration). When spec.gpu is present, this module configures containerd with default_runtime = "nvidia" (via NodeGroupConfiguration) and deploys NFD/GFD/Device Plugin/MIG/DCGM on those nodes.

GPU nodes are usually tainted (for example node-role=gpu:NoSchedule) so regular workloads do not land there. Workloads needing a GPU add matching tolerations and nodeSelector.

Exclusive (one Pod — one GPU)

Each Pod gets an entire physical GPU; the cluster exposes nvidia.com/gpu.

apiVersion: deckhouse.io/v1
kind: NodeGroup
metadata:
  name: gpu-exclusive
spec:
  nodeType: Static # or CloudStatic/CloudEphemeral as needed
  gpu:
    sharing: Exclusive
  nodeTemplate:
    labels:
      node-role/gpu: ""
    taints:
    - key: node-role
      value: gpu
      effect: NoSchedule

TimeSlicing (4 partitions)

Up to four Pods time-share one GPU. Pods still request nvidia.com/gpu.

spec:
  gpu:
    sharing: TimeSlicing
    timeSlicing:
      partitionCount: 4

MIG (profile all-1g.5gb)

Hardware-partition a MIG-capable GPU; resources like nvidia.com/mig-1g.5gb appear.

spec:
  gpu:
    sharing: MIG
    mig:
      partedConfig: all-1g.5gb

For per-index custom MIG partitioning use partedConfig: custom with customConfigs (see FAQ).

Enable DRA mode

Switch the entire GPU management stack to Dynamic Resource Allocation (requires Kubernetes ≥ 1.34):

apiVersion: deckhouse.io/v1alpha1
kind: ModuleConfig
metadata:
  name: gpu
spec:
  enabled: true
  version: 1
  settings:
    dra:
      enabled: true

After applying, the module deploys gpu-controller, gpu-node-agent, and nvidia-adapter into d8-nvidia-gpu. The Device Plugin stack in d8-nvidia-gpu is removed automatically.

Smoke-test Job (CUDA vectoradd)

apiVersion: batch/v1
kind: Job
metadata:
  name: cuda-vectoradd
spec:
  template:
    spec:
      restartPolicy: OnFailure
      nodeSelector:
        node-role/gpu: ""
      tolerations:
      - key: node-role
        value: gpu
        effect: NoSchedule
      containers:
      - name: cuda-vectoradd
        image: nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda11.7.1-ubuntu20.04
        resources:
          limits:
            nvidia.com/gpu: 1

If the Pod finishes with Succeeded, the GPU is present and configured correctly.