nvcf-self-managed-prerequisite

NVCA prerequisites — KAI Scheduler + SMB CSI

One operator tool plus two cluster-level components the NVCA operator / compute plane depends on. Satisfy all three before running nvcf-nvca-install.

The KAI Scheduler and SMB CSI installs are cloud-neutral helm commands pinned to NVCF-validated versions. These are upstream third-party charts (not NVCF images), so they are not in manifest.yaml; the per-component reference docs carry the current pin and link the NVCF docs version table. nvcf-cli is an operator workstation tool, not an in-cluster install.

Prerequisites

• A running Kubernetes cluster (any cloud — AKS, EKS, GKE, k3d, MicroK8s) with kubectl configured and admin access.
• helm >= 3.12 and < 4. Helm 4 is NOT supported (matches nvcf-self-managed-stack/README.md). On Helm 4 the KAI install below hangs silently for many minutes — Helm 4 runs the chart's pre-install crd-manager hook through a before-hook-creation delete and then waits --timeout per already-absent hook resource, so the release sits in pending-install with no pods and never errors cleanly. Use Helm 3.x.
• Cluster has CPU headroom on a general-purpose node pool for KAI's 7 pods.

Install

0 — Preflight: verify Helm 3.x (fail fast on Helm 4)

Helm 4 is not supported and causes a silent multi-minute hang on the KAI install below. Check the major version before installing anything:

helm_major="$(helm version --template '{{.Version}}' | sed -E 's/^v?([0-9]+).*/\1/')"
if [ "

0b — nvcf-cli (compute-plane registration tool)

The compute-plane stack (nvcf-compute-plane-stack) registers each GPU cluster with the control plane via make register-cluster, which shells out to nvcf-cli (init + cluster register) and writes registration/<cluster>-register-values.yaml. The stack's install/apply/sync targets abort if that file is missing. By default the stack builds nvcf-cli from a sibling ../cli checkout (NVCF_CLI_REPO ?= $(MAKEFILE_DIR)/../cli) that is not bundled with the release, so a fresh checkout fails at ensure-nvcf-cli.

nvcf-cli is a convenience wrapper: its only job in this flow is to produce registration/<cluster>-register-values.yaml (the clusterID / selfManaged.clusterId/clusterGroupId schema the NVCA helmfile loads). There are two ways to satisfy this prerequisite.

Register without nvcf-cli (supported path for self-hosted deployments). Obtain the cluster registration data from the running control plane and hand-author registration/<cluster>-register-values.yaml in the schema the NVCA helmfile expects, then run make install CLUSTER_NAME=<name> HELMFILE_ENV=<env> directly (no register-cluster). The step-by-step procedure for gathering that data without the CLI is being published by the NVCF team; until it lands, use an nvcf-cli build if NVIDIA has provided you one.

Use an nvcf-cli binary if you have one. Point the stack at the binary with an absolute path so it does not try to build the missing ../cli:

1 — KAI Scheduler

Skip this step when addons.kaiScheduler.enabled is true in nvcf-compute-plane. That stack installs KAI Scheduler (release and namespace kai-scheduler). Enable that flag whenever Grove or Dynamo is enabled. Use the standalone install below when you need KAI without the compute-plane add-on (for example the NVCA KAIScheduler feature gate alone).

2 — SMB CSI driver

AKS clusters can use the managed csi-driver-smb add-on instead — see references/smb-csi.md.

Definition of done

• Compute-plane registration is satisfiable: either you can produce registration/<cluster>-register-values.yaml without the CLI, or an nvcf-cli binary is available (on PATH or via NVCF_CLI=<abs-path>).
• kubectl get pods -n kai-scheduler shows 7 pods Running.
• kubectl get queues shows both default-parent-queue and default-queue with limit: -1, quota: -1 on cpu / gpu / memory.
• kubectl get csidriver smb.csi.k8s.io returns the driver without error.

After this, run nvcf-nvca-install.

Uninstall

References

• references/kai-scheduler.md — KAI install detail, queue-quota theory, failure modes
• references/smb-csi.md — SMB CSI install detail, AKS managed-add-on alternative, verification
• Companion skill: nvcf-self-managed-installation — Section 7 covers enabling and validating the NVCA operator after these prerequisites are satisfied.