CLI Reference

On this page

Global Flags
l8k discover
l8k generate
l8k deploy
l8k validate
l8k clean
OpenShift Integration
l8k preset list
l8k preset update
l8k sosreport
l8k schema
l8k version
Exit Codes
See Also

The Launch Kit binary is invoked as l8k and exposes a set of subcommands. l8k --help and l8k <subcommand> --help show the complete flag set for the installed version. This page covers the host networking workflow.

Global Flags

Available on all subcommands:

Flag	Description
`--output <string>`	Output format: `text` (default, human-readable) or `json` (structured for automation, see Automation and CI/CD).
`--log-level <string>`	Log level: `trace`, `debug`, `info`, `warn`, `error`.
`--log-file <string>`	Write logs to file instead of stderr.
`-h, --help`	Show help.

--yes and --quiet are root-pipeline flags. Use --output json to auto-confirm lifecycle subcommands. Host workflow commands also accept --config-dir for config/preset overrides and --flavor k8s|ocp for operator integration.

l8k discover

Probes cluster hardware and writes a cluster-config.yaml.

Flag	Description
`--kubeconfig <string>`	Path to kubeconfig file (falls back to `$KUBECONFIG`).
`--user-config <string>`	Base config to merge with discovered hardware. If `--save-cluster-config` is omitted, the file is updated in place.
`--save-cluster-config <string>`	Output path for the discovered cluster configuration. Defaults to `--user-config` if set, otherwise `./cluster-config.yaml`.
`--network-operator-namespace <string>`	Override the Network Operator namespace (default: `nvidia-network-operator`).
`--network-operator-release <string>`	Pin discovery to a Network Operator release line. Supported: `26.1`, `26.4`, `26.7`, `26.10`.
`--spectrum-x <RA-version>`	Enable Spectrum-X for discovery. Accepted values are `RA2.1`, `RA2.2`, `RA2.3`, `RA2.4`, and `RA3.1`; see the version matrix. Required before any of the Spectrum-X flags below — passing one without it is a validation error.
`--multiplane-mode <string>`	Multiplane mode override: `none`, `swplb`, or `hwplb`. Requires `--spectrum-x`.
`--number-of-planes <int>`	Plane count override: `1`, `2`, `4`, or `8` (use RA3.1 VR HWPLB for eight-plane deployments). Requires `--spectrum-x`.
`--topology-scheme <string>`	Topology scheme for Spectrum-X IP allocation: `2-tier` or `3-tier`. Requires `--spectrum-x`. Persisted to `profile.spectrumX.topologyType` in the saved cluster configuration, so a later `l8k generate` does not need the flag.
`--ip-version <string>`	IP version for Spectrum-X address allocation: `ipv4` (default) or `ipv6`. Requires `--spectrum-x`.
`--topology-file <string>`	Path to a spcx-gen/reference-generator or NVIDIA AIR topology JSON, used to generate Spectrum-X `CIDRPool` resources. Requires `--spectrum-x`. Persisted alongside the other Spectrum-X settings.
`--spectrum-x-config <string>`	Spectrum-X input: full doSPCX ConfigMap for RA2.4/RA3.1; full legacy ConfigMap or raw `data.profile` YAML for RA2.3. See Spectrum-X.
`--spectrum-x-configmap-name <string>`	Name for the generated profile ConfigMap. Requires `--spectrum-x`. Needed only when `--spectrum-x-config` holds raw `data.profile` YAML; with a full ConfigMap manifest the name is taken from `metadata.name`.
`--node-selector <string>`	Saved deployment selector; does not limit discovery scheduling. Default: `feature.node.kubernetes.io/pci-15b3.present=true` (Mellanox NICs).
`--image-pull-secrets <strings>`	Image pull secret names for NicClusterPolicy (comma-separated).
`--enabled-plugins <string>`	Comma-separated list of plugins to enable (default: `network-operator`).

l8k generate

Renders Kubernetes manifests from a cluster configuration and a profile selection. Optionally deploys them.

Profile Selection

Flag	Description
`--user-config <string>`	Cluster configuration file. Auto-detected from `./cluster-config.yaml` (or the installed default) if omitted.
`--for <preset>`	Generate for a hardware preset without running discovery. Requires `--node-selector`. See Cluster Topology Presets.
`--fabric <string>`	Fabric type: `ethernet` or `infiniband`. Auto-defaults from the cluster’s unanimous `linkType` when not supplied.
`--deployment-type <string>`	Deployment type: `sriov`, `rdma_shared`, or `host_device`. Auto-defaults to `sriov` when not supplied.
`--multirail`	Enable multirail deployment. Auto-defaults to `true`. Opt out with `--multirail=false` (or set `profile.multirail: false` in YAML).
`--spectrum-x <RA-version>`	Enable Spectrum-X. Accepted values are `RA2.1`, `RA2.2`, `RA2.3`, `RA2.4`, and `RA3.1`; see the version matrix. Implies ethernet fabric, sriov deployment, and multirail.
`--multiplane-mode <string>`	Multiplane mode: `none`, `swplb`, or `hwplb`. Required with `--spectrum-x`; auto-defaulted from the east-west NIC and GPU platform when omitted — ConnectX-7 NIC and BlueField-3 SuperNIC use `none`, ConnectX-9 SuperNIC uses `hwplb`, and ConnectX-8 SuperNIC uses `none` on H100 / H200 / B200 / GB200 and `swplb` on B300 / GB300 or any unrecognised platform. Select `hwplb` explicitly when required by the validated fabric topology. RTX with ConnectX-8 also defaults to `none`. `none` requires `--number-of-planes 1`.
`--number-of-planes <int>`	Number of planes: `1`, `2`, `4`, or `8` (use RA3.1 VR HWPLB for eight-plane deployments). Required with `--spectrum-x`; auto-defaulted from the east-west NIC and GPU platform (ConnectX-7, BlueField-3 SuperNIC, and ConnectX-8 SuperNIC on single-plane platforms → `1`; ConnectX-8 SuperNIC on B300 / GB300 → `2`; ConnectX-9 SuperNIC → `4`, or `8` for VR HWPLB with RA3.1). Pass `4` explicitly for a quad-plane B300 / GB300 topology.
`--topology-scheme <string>`	Topology scheme for Spectrum-X IP allocation: `2-tier` or `3-tier`. Required with `--spectrum-x` — Launch Kit does not default it. May instead be carried in the cluster configuration as `profile.spectrumX.topologyType`, which `l8k discover --topology-scheme` persists; a later `l8k generate` then needs no flag.
`--ip-version <string>`	IP version for Spectrum-X address allocation: `ipv4` (default) or `ipv6`.
`--topology-file <string>`	Path to a spcx-gen/reference-generator or NVIDIA AIR topology JSON. Required with `--spectrum-x` — the Spectrum-X profiles always render `CIDRPool` resources and generation aborts without a topology file. May instead be carried in the cluster configuration as `profile.spectrumX.topologyFile`.
`--spectrum-x-config <string>`	Required profile input for RA2.3/RA2.4/RA3.1. RA2.4/RA3.1 require a full doSPCX ConfigMap and set the NIC template version to the RA; RA2.3 accepts raw legacy YAML and uses the ConfigMap name as its version. See Spectrum-X.
`--spectrum-x-configmap-name <string>`	Name for the generated profile ConfigMap. Required only when `--spectrum-x-config` holds raw `data.profile` YAML; with a full ConfigMap manifest the name is taken from `metadata.name`.
`--network-operator-release <string>`	Pin to a Network Operator release line. Supported: `26.1`, `26.4`, `26.7`, `26.10`. Auto-defaulted under `--spectrum-x`: RA2.4/RA3.1 → 26.10, RA2.3 → 26.7, RA2.2 → 26.4, RA2.1 → 26.1. See Overview.
`--groups <a,b,...>`	Restrict output to the named source groups (comma-separated, matched case-sensitively against `clusterConfig[].identifier`). Mutually exclusive with `--gpu-type`. Empty match is a validation error. See Heterogeneous Clusters.
`--gpu-type <string>`	Restrict output to source groups whose `gpuType` matches (case-insensitive). Mutually exclusive with `--groups`. See Heterogeneous Clusters.
`--node-selector <string>`	Node selector for the synthesized `clusterConfig` when `--for` is used (e.g., `key=value,key2=value2`). Required with `--for`.

Ownership and Routing

--skip-network-operator-helm disables Launch Kit Helm management; external operators remain prerequisites.
--routing destination-based|source-based selects secondary-network routing.
--ignore-arp controls the standard-profile ARP tuning policy.
--overwrite-existing permits preflight convergence, including deleting reported stray CRs. It can affect omitted groups; see Deploy Workflow.

Output and Deployment

Flag	Description
`--save-deployment-files <string>`	Output directory for generated YAML files (default: `./deployment`).
`--deploy`	Apply generated manifests to the cluster.
`--dry-run`	Preview deployment without applying (requires `--deploy`).
`--kubeconfig <string>`	Path to kubeconfig (required with `--deploy`; falls back to `$KUBECONFIG`).

Customization

Flag	Description
`--enable-doca-driver`	Enable DOCA driver deployment.
`--workload-manifest <string>`	Path to a custom workload manifest. Launch Kit patches network annotations, resource requests, and node affinity. See Generate Workflow.
`--image-pull-secrets <strings>`	Image pull secret names for NicClusterPolicy (comma-separated). Propagates to per-group NicNodePolicy sub-specs.
`--network-operator-namespace <string>`	Override the Network Operator namespace.
`--network-namespaces <strings>`	Comma-separated namespaces for secondary-network resources and example workloads.
`--enabled-plugins <string>`	Comma-separated list of plugins to enable (default: `network-operator`).

l8k deploy

Loads the saved bundle, manages Helm when enabled, checks preflight drift, then applies custom resources in four phases: NicClusterPolicy first (await ready), per-group NicNodePolicy (await each), all remaining CRs in one batch (controllers reconcile concurrently), then verify every manifest reached a terminal state. Example workload manifests (*example*) are skipped here — they’re applied by l8k validate --connectivity as part of the data-plane phase. See Deploy Workflow for the full deployment-ordering description.

Flag	Description
`--kubeconfig <string>`	Path to kubeconfig (falls back to `$KUBECONFIG`).
`--deployment-files <string>`	Directory containing the manifests to apply (default: `./deployment`).
`--deploy-timeout <duration>`	End-to-end wall-clock budget for the apply + reconciliation phase (e.g., `45m`, `2h`). `0` (default) means no deadline; polls until every manifest reaches a terminal state. Useful for matching a maintenance window when SR-IOV reconciliation on a large cluster can take an hour or more.
`--dry-run`	Preview what would be applied without changing the cluster.

l8k validate

Checks release and component versions, Helm values when managed, stray resources, certified-preset deviations, manifest readiness, and configured connectivity families. Writes <deployment-files>/k8s-launch-kit-validation-report.html by default. See Validate Workflow for coverage and exit-code semantics.

Flag	Description
`--kubeconfig <string>`	Path to kubeconfig (falls back to `$KUBECONFIG`).
`--user-config <string>`	Explicit source configuration override. When omitted, the bundle sidecar is preferred; see Configuration Reference for lookup order.
`--deployment-files <string>`	Directory containing the manifests to verify (default: `./deployment`).
`--connectivity`	Run the data-plane ping matrix. Default `true`. Pass `--connectivity=false` to limit validate to the static manifest + Helm-release-version checks.
`--connectivity-timeout <duration>`	Wall-clock budget for workload setup and connectivity. Default `0` derives the budget from the matrix plan.
`--validation-checks <strings>`	Select `icmp`, `rping`, and/or `ib_write_bw`.
`--validation-mode <string>`	Coverage and cross-rail expectations: `quick`, `full`, or `strict`.
`--junit-path <string>`	Optional JUnit XML report, independent of HTML output.
`--keep`	Leave the test DaemonSet running after `--connectivity` completes (useful for follow-up debugging).
`--wait <duration>`	Block validate up to this duration waiting for in-progress manifests to reach a terminal state. `0` (default) returns immediately on the first snapshot. Re-polls the cluster every 10 s.
`--report-path <string>`	Write the HTML report to this path. Empty (default) writes to `<deployment-files>/k8s-launch-kit-validation-report.html`. Pass `-` to skip the report file entirely.

l8k clean

Removes Network Operator resources and, by default, its Helm release. This is a cluster mutation with no dry-run mode. Inspect the target kubeconfig and namespace first. --keep-helm-chart retains Helm; networkOperator.skipHelmChart also preserves externally managed Helm. The namespace and CRDs are preserved. OpenShift cleanup is unsupported. See l8k clean --help for the current removal scope and flags.

OpenShift Integration

Use --flavor ocp or flavor: ocp with separately installed certified Network Operator, SR-IOV Operator, Maintenance Operator, and NFD prerequisites. Multus and reference CNI plugins remain platform-owned. See Getting Started with OpenShift and the Launch Kit OpenShift guide for ownership and namespace configuration.

l8k preset list

Lists locally available cluster topology presets.

l8k preset list

l8k preset update

Downloads the latest presets from a Git repository.

Flag	Description
`--repo <string>`	GitHub repository in `owner/repository` form.
`--branch <string>`	Branch or tag to fetch from (default: `main`).
`--dir <string>`	Local destination directory for downloaded presets.

See Cluster Topology Presets for usage.

l8k sosreport

Collects diagnostic data from the cluster.

Flag	Description
`--kubeconfig <string>`	Path to kubeconfig (falls back to `$KUBECONFIG`).
`--output-dir <string>`	Directory to save the sosreport (default: `./sosreport`).

See Troubleshooting for analysis guidance.

l8k schema

Outputs a JSON description of l8k’s capabilities — supported commands, fabrics, deployment types, exit codes, and output formats. Used by AI agents and automation tooling to discover l8k capabilities programmatically. See Automation and CI/CD.

l8k schema

l8k version

Prints the version. Supports --output json for structured output.

Exit Codes

Code	Meaning
0	Success.
1	General error.
2	Validation error (invalid flags or configuration).
3	Cluster error (API unreachable, discovery failed).
4	Deployment or validation gate failure.
5	Partial success (discovery completed but deployment failed).