User Guide

End-to-End Tutorial

View as Markdown

This tutorial walks the full AICR workflow once, end to end: install the CLI, generate a recipe for your environment, render it into deployment bundles, deploy them, and validate the running cluster against the recipe. It is a learning path — follow it top to bottom on a non-production cluster to build a mental model of how the four stages fit together.

For the conceptual overview of the four stages, see the documentation hub. For exhaustive flag lists, see the CLI Reference.

Prerequisites

  • A GPU-accelerated Kubernetes cluster you can deploy to (EKS, GKE, AKS, or a local Kind/KWOK cluster for a dry run). kubectl configured to reach it.
  • The helm binary on your PATH (the default helm deployer emits Helm commands).
  • About 15 minutes. No NVIDIA hardware is required to generate a recipe or a bundle — only the deploy and validate stages touch a real cluster.

Step 1 — Install the CLI

$# Homebrew
$brew tap NVIDIA/aicr
$brew install aicr
$
$# Or the install script
$curl -sfL https://get.aicr.run | bash -s --
$
$aicr --version

For manual installation, container images, or building from source, see Installation.

Step 2 — Generate a recipe

A recipe is a version-locked configuration for a specific environment. Describe your target with criteria flags and AICR matches it against its library of validated overlays:

$aicr recipe \
>  --service eks \
>  --accelerator h100 \
>  --os ubuntu \
>  --intent training \
>  --platform kubeflow \
>  --output recipe.yaml

Open recipe.yaml — it lists the components that will be deployed, their pinned versions, the declarative constraints, and the deployment order. The valid values for each criterion (services, accelerators, operating systems, intents, platforms) are enumerated in the CLI Reference and the documentation hub glossary.

Prefer to start from your live cluster instead of criteria? Capture a snapshot first (aicr snapshot --output snapshot.yaml) and pass --snapshot snapshot.yaml to aicr recipe. See Agent Deployment for in-cluster snapshot capture.

Step 3 — Inspect a resolved value (optional)

Before bundling, you can query any hydrated value without rendering the whole bundle — useful for scripting and sanity checks:

$aicr query \
>  --service eks --accelerator h100 --os ubuntu --intent training --platform kubeflow \
>  --selector components.gpu-operator.values.driver.version

Step 4 — Render deployment bundles

The bundler materializes the recipe into deployment-ready artifacts — one folder per component, each with Helm values, checksums, and a README:

$aicr bundle --recipe recipe.yaml --output ./bundles

With the default helm deployer, ./bundles contains per-component folders and a deploy.sh that runs the Helm installs in dependency order. To target a GitOps tool instead (Argo CD, Flux, Helmfile), or to override values and scheduling, see Generating Bundles.

Step 5 — Deploy to your cluster

$cd bundles
$chmod +x deploy.sh
$./deploy.sh

This installs each component in order. Watch the GPU Operator and any platform components come up with kubectl get pods -A -w.

Step 6 — Validate the running cluster

The validator compares the recipe against the live cluster — first the declarative constraints, then optional in-cluster phases (deployment, performance, conformance):

$aicr validate --recipe recipe.yaml

A clean run exits 0. For the phase model, performance testing, and emitting signed evidence for a recipe PR, see Validation.

Where to go next