Artifact Manifest

This section provides a comprehensive list of all components required for NVIDIA Cloud Functions (NVCF) Self-Hosted deployment for basic inference. Additional components are needed for Low Latency Streaming (Simulation).

Artifacts Overview

The following inventories list the artifacts for an inference-only self-hosted NVCF deployment. Artifacts are grouped by deployment plane and type.

Prepare Helm charts for Helmfile

The self-managed Helmfile bundles currently expect NVCF charts in an OCI registry. Chart distributions that start with https:// are Helm repository charts. Before deploying with Helmfile, copy each required repository chart version into the OCI registry configured by global.helm.sources in the Helmfile bundles.

The following example copies one public NVCF chart into an OCI registry:

$
export CHART_NAME="helm-nvcf-api"
$
export CHART_VERSION="1.22.5"
$
export TARGET_REGISTRY="<registry-host>"
$
export TARGET_REPOSITORY="<repository>"
$
$
helm repo add nvcf https://helm.ngc.nvidia.com/nvidia/nvcf
$
helm repo update
$
helm pull "nvcf/${CHART_NAME}" --version "${CHART_VERSION}"
$
$
helm registry login "${TARGET_REGISTRY}"
$
helm push "${CHART_NAME}-${CHART_VERSION}.tgz" \
>
  "oci://${TARGET_REGISTRY}/${TARGET_REPOSITORY}"

Repeat this process for every required chart with an https:// distribution. Copy required charts with an nvcr.io distribution into the same target repository so Helmfile can resolve all NVCF charts from one source. Configure the stack environment with that OCI location:

1
global:
2
  helm:
3
    sources:
4
      registry: "<registry-host>"
5
      repository: "<repository>"

See Image Mirroring for additional registry examples.

Supporting images that ship from Docker Hub

The NATS configuration reloader and the account-bootstrap Kubernetes utilities are not republished under the public nvidia/nvcf catalog, so they default to their upstream Docker Hub source:

A public-catalog install needs no configuration for these two images. Verify that your cluster can reach Docker Hub. If your egress to Docker Hub is authenticated or rate-limited, add the pull secret to global.imagePullSecrets.

If you have mirrored both images into your own registry, redirect them from your environment file (deploy/stacks/self-managed/environments/<environment>.yaml). Set registry and repository together, because each key falls back to its own Docker Hub default rather than to global.image. You do not need to edit deploy/stacks/self-managed/global.yaml.gotmpl.

1
nats:
2
  reloader:
3
    image:
4
      registry: <your-registry>
5
      repository: <your-repository>/nats-server-config-reloader
6
      tag: "0.23.0"
7
8
api:
9
  accountBootstrap:
10
    image:
11
      registry: <your-registry>
12
      repository: <your-repository>/alpine-k8s
13
      tag: "1.36.1"

Override the Cassandra images

The Cassandra server and its dynamic seed discovery container take the same override keys, but they default to global.image. Set them when your registry uses a different path or tag than the cassandra default under global.image.repository:

1
cassandra:
2
  image:
3
    registry: nvcr.io
4
    repository: nvidia/nvcf/cassandra
5
  dynamicSeedDiscovery:
6
    image:
7
      registry: nvcr.io
8
      repository: nvidia/nvcf/cassandra

Every registry, repository, and tag key is optional. Any key you leave unset keeps its default: for Cassandra, registry and repository fall back to global.image.registry and global.image.repository; for the reloader and account-bootstrap images they fall back to the Docker Hub coordinates above. tag falls back to the version the stack pins, or to the chart’s own default when the stack pins none.

Use the version listed in the artifact table. Verify that your cluster can reach the registry you point each image at. If it requires authentication, add its pull secret to global.imagePullSecrets.

The current Cassandra initialization hook uses the nvcf-cassandra-migrations image, and the current NATS chart renders NKeys as Secrets without an nkey job. Their legacy cassandra.initialization.image and nats.nkeyJob.image values do not control rendered workloads. Using alpine-k8s for those operations requires chart support rather than a configuration-only override.

The following tables list the complete artifact inventory.

Control plane Helm charts

Control plane services and images

Compute plane Helm charts

Compute plane services and images

EA-only CVE-impacted artifacts

These Early Access artifacts have known CVE impact. Use only the QA-qualified versions listed for this EA stack.

Tools and deployment resources