API reference

All types belong to API group resource.nvidia.com/v1beta1.

GPU opaque configuration types

These types are set as opaque configuration in ResourceClaim and ResourceClaimTemplate specs to configure GPU resources. They are optional, omit them to use driver defaults. Some fields required feature gates to be enabled before using.

GpuConfig

Configures a full GPU device. Target DeviceClass: gpu.nvidia.com.

With time-slicing:

apiVersion: resource.nvidia.com/v1beta1
kind: GpuConfig
sharing:
  strategy: TimeSlicing
  timeSlicingConfig:
    interval: Default       # Default | Short | Medium | Long

With MPS (Multi-Process Service):

apiVersion: resource.nvidia.com/v1beta1
kind: GpuConfig
sharing:
  strategy: MPS
  mpsConfig:
    defaultActiveThreadPercentage: 50             # optional, integer
    defaultPinnedDeviceMemoryLimit: "4Gi"         # optional, quantity
    defaultPerDevicePinnedMemoryLimit:            # optional, map
      "0": "2Gi"

Fields

MigDeviceConfig

Configures a MIG device slice. Target DeviceClass: mig.nvidia.com.

With time-slicing:

apiVersion: resource.nvidia.com/v1beta1
kind: MigDeviceConfig
sharing:
  strategy: TimeSlicing

With MPS:

apiVersion: resource.nvidia.com/v1beta1
kind: MigDeviceConfig
sharing:
  strategy: MPS
  mpsConfig:
    defaultActiveThreadPercentage: 50

Fields

Note: timeSlicingConfig is not available for MIG devices. Time-slicing on MIG slices does not support interval configuration.

VfioDeviceConfig

Configures a VFIO passthrough device. Target DeviceClass: vfio.gpu.nvidia.com.

apiVersion: resource.nvidia.com/v1beta1
kind: VfioDeviceConfig
iommu:
  backendPolicy: LegacyOnly   # LegacyOnly | PreferIommuFD
  enableAPIDevice: false       # optional

Fields

Requires PassthroughSupport (Alpha, default: false). iommu.enableAPIDevice additionally requires DeviceMetadata (Alpha, default: false).

ComputeDomain CRDs

ComputeDomains use two Custom Resource Definitions. Unlike the GPU opaque types above, these are concrete Kubernetes resources that exist independently of ResourceClaim specs.

ComputeDomain

A user-created resource that provisions an ephemeral multi-node NVLink fabric. Creating one triggers the controller to deploy a per-domain IMEX daemon fleet and generate ResourceClaimTemplate objects that workload pods reference to claim IMEX channels. The spec is immutable after creation.

apiVersion: resource.nvidia.com/v1beta1
kind: ComputeDomain
metadata:
  name: my-compute-domain
spec:
  numNodes: 0
  channel:
    resourceClaimTemplate:
      name: my-channel
    allocationMode: Single        # Single (default) | All

Spec fields

Status fields

Note: Do not gate workload startup on status.status. The status field is informational. IMEX daemons start as soon as their local node joins without waiting for all peers.

ComputeDomainClique

A ComputeDomainClique is a driver-managed resource created automatically by the IMEX daemon on each node. You do not create or modify these directly.

Each ComputeDomainClique is namespaced to the driver namespace and named as <computeDomainUID>.<cliqueID>. It tracks which nodes have joined a specific NVLink clique within a ComputeDomain and their daemon readiness.

The ComputeDomain plugin reads these objects to verify that the local IMEX daemon is ready before allowing a workload container to start.

Requires ComputeDomainCliques (Beta, default: true), which depends on IMEXDaemonsWithDNSNames.

Fields

ComputeDomain opaque configuration types

The following opaque types are set in the ResourceClaim specs generated by the ComputeDomain controller. They are managed automatically, you do not set these directly. They are documented here for reference and debugging.

ComputeDomainChannelConfig

Opaque configuration for IMEX channel ResourceClaim objects. Set by the controller in the ResourceClaimTemplate it generates from spec.channel.resourceClaimTemplate.

ComputeDomainDaemonConfig

Opaque configuration for IMEX daemon ResourceClaim objects. Set by the controller in the daemon DaemonSet.