Skip to navigation

NVIDIA Run:ai supports a broad range of workloads from the ML and Kubernetes ecosystem that are already registered as workload types in the platform and ready to use. These workloads are managed by Karta, which ensures they receive the same advanced scheduling, orchestration, and monitoring capabilities as NVIDIA Run:ai native workloads. For more details on feature support, see Supported features.

All supported workload types are pre-registered using the Workload Types API. You can retrieve the full list of these workloads and their current configuration defaults.

Note

To define and register workload types beyond those provided by NVIDIA Run:ai, see Extending workload support with Karta.

List of Supported Workload Types

The following workload types are already registered in the platform and ready to use:

Scope of Slinky NodeSet Support

NVIDIA Run:ai recognizes a Slinky NodeSet - the group of Slurm compute node (slurmd) pods that the Slinky operator runs on Kubernetes - and manages it as a standard workload. The NodeSet receives the same scheduling, prioritization, preemption, and monitoring as any other supported workload type, and appears in the Workloads table for management.

This support is scoped to the NodeSet workload as it runs on Kubernetes. It is not an integration with the Slurm control plane: NVIDIA Run:ai does not submit, schedule, or manage Slurm jobs, partitions, accounts, or the Slurm controller (slurmctld).

Submitting Supported Workload Types

Supported workload types are submitted as standard Kubernetes YAML manifests**.** Once submitted, the workload is created and appears in the Workloads table for monitoring and management. See Submit supported workload types via YAML.

Workload Types and Defaults

Each workload type is assigned a default category, which determines its default priority and preemptibility. These defaults influence how workloads are scheduled and prioritized within a project, as well as how they are grouped for monitoring and reporting.

Note

To retrieve the default category, priority and preemptibility per workload type, refer to the List workload types API.

CategoryDefault PriorityDefault Preemptibility
BuildHighNon-preemptible
TrainLowPreemptible
DeployVery highNon-preemptible

Workload Types by Default Category

Workload TypeBuildTrainDeploy
AMLJob✗✓✗
CronJob✗✗✓
Deployment✗✗✓
DevWorkspace✓✗✗
DynamoGraphDeployment✗✗✓
InferenceService (KServe)✗✗✓
JAXJob✗✓✗
Job✓✗✗
JobSet✗✓✗
LeaderWorkerSet (LWS)✗✗✓
MPIJob✗✓✗
NIMCache✗✗✓
NIMServices✗✗✓
NodeSet (Slinky)✗✗✓
Notebook✓✗✗
PipelineRun✗✗✓
Pod✗✗✓
PyTorchJob✗✓✗
RayCluster✗✓✗
RayJob✗✓✗
RayService✗✗✓
ReplicaSet✗✗✓
ScheduledWorkflow✗✗✓
SeldonDeployment✗✗✓
Service✗✗✓
SPOTRequest✗✗✗
StatefulSet✗✗✓
TaskRun✓✗✗
TFJob✗✓✗
VirtualMachineInstance✗✓✗
Workflow✗✗✓
XGBoostJob✗✓✗

Updating Default Category and Priority Mapping

Administrators can change the default priority and category assigned to a workload type by updating the mapping using the NVIDIA Run:ai API: