Overview
NVIDIA Kubernetes Launch Kit¶
NVIDIA Kubernetes Launch Kit (l8k) helps platform teams configure NVIDIA networking for accelerated Kubernetes clusters. It discovers NIC and GPU topology, generates reviewable Network Operator and NIC Configuration Operator resources, applies them in dependency order, and checks the deployed result.
Use this site when you are deploying SR-IOV, RDMA shared-device, host-device, InfiniBand, or Spectrum-X networking. Start with Plan your deployment to choose a platform and profile, check what is qualified, and identify the site inputs and permissions you need.
The lifecycle commands operate on the host target by default. For a separate
DPU estate, the dpf target supports standalone inventory discovery,
HBN-BGP artifact generation, and
deployment with provisioning and runtime checks. The
current BF4/HBN-BGP cohort remains beta; command availability does not qualify
every hardware and release combination.
Management probes cover every observed node that meets the existing readiness and
platform checks; node count alone does not establish control-plane HA or production
qualification. DPF standalone validation and the root
pipeline remain unavailable. See the DPF workflow.
Find Your Path¶
| If you are a... | Start here |
|---|---|
| Operator preparing a deployment | Plan your deployment |
| Operator deploying SR-IOV Ethernet on Kubernetes | Quick Start |
| OpenShift administrator | Deploy on OpenShift |
| Operator inventorying a cluster | Cluster Discovery |
| Platform engineer selecting a topology profile | Deployment Profiles |
| Platform engineer managing mixed hardware | Heterogeneous Clusters |
| Spectrum-X operator | Spectrum-X |
| CI/CD or GitOps integrator | Automation |
| AI agent integrator | AI Skills |
| Operator confirming a deployment is ready for use | Validation |
| Application owner connecting a workload | Use the network in an application |
| Operator changing an existing deployment | Change and upgrade |
| Operator removing a deployment | Cleanup |
| Operator investigating a failed stage | Troubleshooting |
Workflow¶
Each stage is independently invocable:
l8k discoverbootstraps a private NIC discovery daemon and writescluster-config.yaml.l8k generaterenders a profile-specific manifest bundle underdeployment/network-operator/.l8k deployinstalls or upgrades the Network Operator Helm chart, applies CRs in dependency order, and waits for reconciliation.l8k validateruns deployment checks and produces a report. Review its completed checks and connectivity coverage before acceptance; exit success alone is insufficient.l8k cleanremoves Network Operator custom resources and uninstalls its Helm release unless config marks the chart externally owned.
Review discovered inventory, site networking settings, and generated artifacts before deploying. Quick Start shows that checkpoint in one Kubernetes scenario. A successful validation exit alone does not demonstrate complete data-plane coverage; use the acceptance criteria.