DocumentationSelf-Managed DeploymentSetupHelm

Prerequisites

View as Markdown

Before installing the NeMo Platform Helm chart, prepare Kubernetes access, storage, and NGC credentials.

Kubernetes cluster access

Install and configure these tools on the machine where you run the install:

  • kubectl configured for the target cluster
  • helm 3.x
  • Permission to create namespaces, secrets, RBAC resources, services, deployments, persistent volume claims, and cluster-scoped resources required by chart dependencies

Set a namespace variable used by the examples:

1export NMP_NAMESPACE=nemo-platform
2kubectl create namespace "$NMP_NAMESPACE"

Review the system requirements before choosing node types. For storage and persistent volume configuration, including ReadWriteMany and StorageClass requirements, see Persistent Volumes.

Cluster requirements

The chart can run on local or managed Kubernetes clusters when the required storage, networking, and image access are available.

ClusterMinimum setup
minikubeSingle-node smoke testing. Enable or provide a StorageClass and use port-forwarding or the ingress addon for API access.
kindSingle-node smoke testing. Provide a static PersistentVolume or StorageClass for the shared PVC and use port-forwarding for API access.
EKSConfigure EFS or another RWX storage backend for shared jobs storage. Use GPU node groups and NVIDIA GPU Operator or device plugin for GPU workloads. Configure EFA only for multi-node GPU jobs.
AKS, GKE, OKE, or on-premConfigure an RWX-capable StorageClass, GPU node support, ingress or Gateway API, and any cloud-specific high-performance networking before enabling multi-node jobs.
OpenShiftApply the OpenShift security context overrides in OpenShift.

Create NGC API key and secrets

You need an NGC account and an NGC API key that can access the NeMo Platform chart and container images in the nvidia/nemo-platform org. NeMo Platform images are hosted under paths such as nvcr.io/nvidia/nemo-platform/nmp-api and nvcr.io/nvidia/nemo-platform/safe-synthesizer-tasks.

To learn more about using Helm charts in the NGC Catalog console, refer to the Helm Charts section in the NGC Catalog User Guide.

  1. Create an NGC API key following Generating NGC API Keys.

    Include permissions that allow access to NGC Catalog artifacts and the NVIDIA private registry. At minimum, include the NGC Catalog and Public API Endpoints permissions shown in the NGC API key UI. If your organization uses scoped service keys, ensure the key can read the nvidia/nemo-platform org.

  2. Store the key in your shell environment.

    1export NGC_API_KEY=<your-ngc-api-key>

  3. Create the image pull secret.

    This secret lets Kubernetes pull container images from nvcr.io.

    1kubectl create secret docker-registry nvcrimagepullsecret \
    2  --docker-server=nvcr.io \
    3  --docker-username='$oauthtoken' \
    4  --docker-password="$NGC_API_KEY" \
    5  --namespace "$NMP_NAMESPACE"

  4. Create the NGC API secret.

    This secret is mounted into platform services that need to authenticate to NGC APIs for model and artifact access.

    1kubectl create secret generic ngc-api \
    2  --from-literal=NGC_API_KEY="$NGC_API_KEY" \
    3  --namespace "$NMP_NAMESPACE"

  5. Confirm both secrets exist.

    1kubectl get secret nvcrimagepullsecret ngc-api -n "$NMP_NAMESPACE"

For more information about other secrets you might configure through the Helm chart, see Manage Secrets.