Skip to navigation
Development

External Dependencies

View as Markdown

NVIDIA Run:ai’s cluster components build on external open-source projects, pulled in as versioned Helm chart dependencies or Go modules rather than built from Run:ai source. NVIDIA Run:ai packages, configures, and manages these dependencies as part of the platform, so no separate installation or maintenance is required to use them.

Helm chart dependencies are declared in runai-chart/Chart.yaml and documented in the Chart dependencies section of the runai-chart README.

KAI Resource Management (KRM)

Source: kai-scheduler/kai-resource-management Dependency of: runai-chart - a direct Helm chart dependency

KAI Resource Management is a Kubernetes-native resource-management layer for KAI Scheduler, adding organizational abstractions - projects, departments, node pools, queues - and workload placement. Its Helm chart packages nodepool-controller, project-controller, pod-group-assigner, and krm-operator, replacing the versions of those controllers previously built from Run:ai’s own cluster source, and declares KAI Scheduler as its own sub-chart dependency.

KAI Scheduler

Source: kai-scheduler/KAI-Scheduler Dependency of: KRM - a sub-chart dependency (oci://ghcr.io/kai-scheduler/kai-scheduler, declared in KRM’s Chart.yaml), not a direct entry in runai-chart

KAI (Kubernetes AI) Scheduler is an open-source, CNCF-donated AI-aware scheduling engine. It applies fair-share, preemption, bin-packing, and multi-node placement rules to GPU workloads, and is the engine underlying the NVIDIA Run:ai Scheduler.

NVIDIA Run:ai integrates KAI with specific configuration overrides - the go-operator owns kai-config and the kai-nodepool-controller creates one KAI SchedulingShard per node pool. KAI’s default queue, shard, and config-deployer components are disabled.

gpu-fractioning

Source: kai-scheduler/gpu-fractioning Dependency of: KAI Scheduler - a sub-chart dependency, so it does not need its own entry in runai-chart/Chart.yaml

gpu-fractioning is a Kubernetes operator that enables multiple pods to safely share a single GPU with enforced memory boundaries. It runs alongside KAI Scheduler: KAI Scheduler decides which fraction of which GPU a workload gets, and gpu-fractioning enforces that memory boundary on the node and exports per-pod metrics for the resulting fractional GPUs.

kai-resource-management-api

Source: kai-scheduler/kai-resource-management-api Dependency of: Run:ai’s own cluster services (declared in each service’s go.mod) - a Go module dependency, not a Helm chart dependency and not a separately running service

The API-only counterpart to KRM: the CRD type definitions (kai.resources: Project, Department, NodePool, ManagedNodesConfig, KRMConfig). Any Run:ai service that reads or writes these CRDs imports this module for the typed Go structs, independent of which controller owns reconciling them - today that’s a mix of Run:ai’s own controllers and, once KRM is adopted, KRM’s. Migrating legacy CRDs onto kai.resources runs automatically this release and self-gates once complete; it is not a customer-facing toggle.

workload-map

Source: dsx-ai-factory/workload-map Dependency of: runai-chart - a direct Helm chart dependency (oci://ghcr.io/dsx-ai-factory/workload-map/karta, declared in Chart.yaml)

Karta is a CRD-based resource abstraction layer with its own operator. It replaced the legacy ResourceInterface (RI) CRD as the way Run:ai represents and reconciles workload resource attachments in the cluster. The karta-migrate Helm hook handles the one-time migration from RI to Karta CRs on upgrade.

Integrating a Custom Version

To replace any of these dependencies with a custom build:

  1. Fork the relevant external repository and make your changes.
  2. Build and publish your custom chart to an OCI registry.
  3. Update the dependency entry in runai-chart/Chart.yaml - change repository to your registry and version to your build.
  4. Run helm dep update in runai-cluster/cluster/runai-chart/ to pull in your version.
  5. Deploy with helm upgrade or via local-dev.sh.