k8s-launch-kit-shared

par nvidia

CLI k8s-launch-kit (l8k) : Modèles partagés pour l'emplacement des binaires, les indicateurs globaux, le formatage de la sortie, les codes de sortie et la gestion des erreurs. Lisez ceci avant d'utiliser tout…

npx skills add https://github.com/nvidia/k8s-launch-kit --skill k8s-launch-kit-shared

l8k — Shared Reference

Installation

# Build and install (production — copies binary + profiles)
make build && sudo scripts/install.sh

# Development (symlinks into source tree)
make build && scripts/install.sh --dev-env

Install Paths

PathContents
/usr/local/bin/l8kCLI binary (on PATH)
/usr/local/share/l8k/profiles/Go template profiles
/usr/local/share/l8k/presets/Topology presets (per (machineType, gpuType) directories)
/usr/local/share/l8k/l8k-config.yamlDefault config

After installation, l8k is available system-wide.

Binary Discovery (for AI agents)

CRITICAL — follow these steps exactly:

  1. Run which l8k in a single Bash call. If it returns a path, use it. Done.
  2. If which l8k fails (exit code 1), tell the user l8k is not installed. Point them to: make build && sudo scripts/install.sh. Stop — do not proceed.
  3. NEVER do any of the following to find l8k:
    • find or ls commands
    • Spawn a subagent to locate the binary
    • Check multiple paths or directories
    • Explore the repo contents

Available Commands

CommandDescription
l8k discoverDiscover cluster hardware and produce cluster-config.yaml
l8k generateGenerate Kubernetes YAML manifests from config + profile (use --for <preset> to skip cluster discovery for known SKUs)
l8k deployApply generated manifests and install or upgrade the Network Operator Helm release
l8k cleanDelete Network Operator custom resources and optionally uninstall its Helm release
l8k validateVerify the Helm release, manifests, component versions, and connectivity
l8k preset listList bundled topology presets (directory + machineType + gpuType)
l8k preset updateDownload latest topology presets from GitHub
l8k sosreportCollect diagnostic dump from cluster
l8k schemaList all capabilities as JSON (profiles, flags, exit codes)
l8k versionPrint version information

The root command l8k --discover-cluster-config ... still works for backward-compatible full-pipeline usage.

Target selection

discover, generate, deploy, validate, and the root pipeline default to the host target. Existing invocations must omit --target unless the user explicitly requests it; --target host is an equivalent explicit form.

The dpf target name is reserved but its phases are not implemented in this build. Do not attempt DPF provisioning. Use l8k schema and inspect targets[].phases before selecting any non-host target. Host-only flags such as --fabric, --kubeconfig, and the Network Operator/Spectrum-X flags are rejected when explicitly combined with another target.

Internally, each lifecycle command snapshots the explicitly supplied Host arguments, binds a phase-specific adapter through the target registry, and runs the resulting operation exactly once with the command context. Do not bypass this route when extending or automating a lifecycle phase. The canonical component and data-flow diagrams live in docs/architecture/overview.md; update them with any change to lifecycle, package ownership, artifacts, or external integration boundaries.

Global Flags

FlagDescription
--target <NAME>Lifecycle target: host (default); dpf is currently unavailable
--kubeconfig <PATH>Path to kubeconfig file (optional — falls back to $KUBECONFIG env var)
--user-config <PATH>Path to user-supplied l8k-config.yaml
--output <FORMAT>Output format: text (default), json
--yes / -yAuto-confirm all prompts (root command only — not available on subcommands; --output json auto-confirms)
--quiet / -qSuppress informational output (errors still shown)
--log-level <LEVEL>Enable logs at trace, debug, info, warn, or error. Debug shows structured progress; trace also shows bounded command output.
--network-operator-namespace <NS>Override network operator namespace (default: nvidia-network-operator). No-op for l8k discover — discover always bootstraps into nvidia-k8s-launch-kit; the flag still applies to l8k generate / l8k deploy / l8k clean / l8k validate.
--network-namespaces <NS,...>Comma-separated namespaces for the secondary-network CRs + example test DaemonSets; one copy rendered per namespace (shared resources like IPPools/NodePolicies are NOT duplicated). Default: default
--node-selector <LABELS>Restrict to nodes matching labels (comma-separated, ANDed)
--image-pull-secrets <NAMES>Image pull secret names for Network Operator components and authenticated Helm chart downloads (comma-separated)

l8k discover and l8k generate both accept the profile flags --fabric, --deployment-type, --multirail, --spectrum-x, --multiplane-mode, and --number-of-planes. Discovery persists the resolved values; later generation reuses them unless another explicit CLI override is supplied.

Agent / JSON Mode

RULE: AI agents MUST always use --output json 2>/dev/null when calling any l8k subcommand. Never use text mode — it produces unstructured output with spinners and ANSI codes that is hard to parse.

Do NOT use --yes with subcommands — it only exists on the root command and will cause "unknown flag" errors. --output json already auto-confirms all prompts (no interactive input needed).

# Correct
l8k discover --output json 2>/dev/null
l8k generate --output json 2>/dev/null

# WRONG — --yes is not a valid flag on subcommands
l8k discover --output json --yes 2>/dev/null
  • stdout: Exactly one JSON object (JSONResult) at completion
  • stderr: Human-readable log lines
  • Pipe with jq for downstream processing: l8k discover ... --output json 2>/dev/null | jq .success

Exit Codes

CodeMeaning
0Success
1General/unknown error
2Validation error — bad flags, missing required arguments, or invalid config
3Cluster error — kubeconfig invalid, API unreachable, missing CRDs, no NVIDIA NICs
4Deployment error — apply failed
5Partial success — discovery ok but deploy failed

Structured Error Output (JSON mode)

{
  "success": false,
  "phase": "discover",
  "deployed": false,
  "error": {
    "code": "CLUSTER_ERROR",
    "message": "cluster discovery failed",
    "category": "cluster",
    "transient": true,
    "suggestion": "Check that kubeconfig is valid and the cluster is reachable"
  },
  "messages": [...]
}

Error categories: validation, cluster, deployment. The transient field hints whether retrying might help.

Schema Discovery

# List all l8k capabilities as JSON
l8k schema

Use l8k schema to programmatically discover available profiles, fabrics, deployment types, flags, exit codes, and output formats.

Security Rules

  • Confirm with the user before executing --deploy on a production cluster
  • Prefer --dry-run for destructive operations
  • Never expose kubeconfig contents in output

Network Operator Namespace Resolution

Applies to l8k generate / l8k deploy / l8k validate only. l8k discover manages its own private namespace (nvidia-k8s-launch-kit) and ignores this flag.

Both nvidia-network-operator and network-operator are common default namespaces for an existing Network Operator install. If l8k generate / l8k deploy / l8k validate can't find Network Operator resources, retry with --network-operator-namespace <correct-namespace>.

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…