AI Runway AKS Deployment Guide
Turn an existing AKS cluster into a validated environment for running AI models.
The documentation requires confirmation before installation or deployment, warns about GPU costs, and uses interactive HuggingFace token handling with temporary-file cleanup. However, kubectl/make operations can materially change the cluster, while least-privilege scope, complete data-flow disclosure, systematic rollback, and recovery after failed installation are not defined, so points were deducted.
The sequence, state reporting, prerequisite checks, and several troubleshooting paths are reasonably clear. However, the referenced step-6-summary.md and troubleshooting.md are not supplied, commands depend on an external repository and unspecified versions, and skip-to-step assumes prior completion; static evidence cannot establish reproducible key paths, so points were deducted.
Triggers, audience, and CPU/GPU plus KAITO/Dynamo/KubeRay scenarios are fairly explicit. Boundaries for unsupported AKS configurations are limited, Chinese-language interaction is not addressed, and reachability of HuggingFace, container images, and model downloads from mainland China is unspecified, so points were deducted.
The skill uses layered step references, includes MIT, author, and version metadata, and provides naming rules, model examples, and some PowerShell variants. It lacks a changelog, explicit maintenance owner and update path; some referenced files are not supplied, and limitation/troubleshooting coverage is incomplete, so points were deducted.
It covers the intended onboarding flow from cluster verification through controller, provider, and first-model deployment, so the core task is plausibly supported. But there is no skill-specific verification evidence, rollback outcome, or executed result; some model recommendations appear mismatched or require additional configuration, leaving substantial field judgment and rework, so points were deducted.
The supplied source is auditable, and the repository includes general CI and test-harness material. However, the shown tests do not cover this AKS skill path, no skill-specific acceptance scenarios or third-party execution evidence are provided, and there is no corroboration, so only limited credit is justified.
- Before execution, confirm the active kubeconfig context, target namespace, RBAC permissions, and billing impact; a general rollback procedure is not provided.
- Do not rely on skip-to-step without independently checking state; the option explicitly assumes earlier steps are complete.
- Model, provider, GPU, image, and HuggingFace downloads require additional version and network validation; mainland-China reachability is unspecified.
- Verify that step-6-summary.md, troubleshooting.md, the AI Runway repository root, and provider Makefiles actually exist and match the documented commands.
What it does & when to use it
This skill is for users who already have an AKS cluster and want to deploy AI Runway. It walks through cluster verification, controller and CRD installation, GPU assessment, inference-provider setup, and a first model deployment. Users can resume from a numbered step, and the skill requires confirmation before installation or deployment actions. It also calls out the potentially significant cost of GPU node pools.
Checks the Kubernetes context, node inventory, and GPU availability; runs cluster operations with kubectl, make, and curl; installs the AI Runway controller and CRDs; assesses GPU models and dtype or attention constraints; recommends and installs an inference provider such as KAITO, Dynamo, or KubeRay; deploys an initial model and verifies Ready status; reports each step as healthy or failed and provides commands for common failure modes.
- A cloud engineer has an AKS cluster and needs to install the AI Runway controller from scratch.
- An ML platform team needs to check whether its AKS GPUs are suitable for model inference.
- A model-serving developer wants to deploy a first model through KAITO, Dynamo, or KubeRay.
- A user with a partially completed AI Runway setup needs to resume from a specific phase.
- An AKS operator is troubleshooting GPU scheduling, controller failures, or bfloat16 inference errors.
Pros & cons
- Covers the full path from cluster checks to first-model verification.
- Requires explicit status reporting and skips steps that are already complete.
- Addresses GPU compatibility, provider selection, and common failure modes.
- Uses no MCP server; operations are performed directly through kubectl and make.
- Requires an existing AKS cluster; cluster provisioning is out of scope.
- GPU pools can be expensive; the source cites A100-80GB costs of $3–5 or more per hour.
- The detailed step references are not included in the supplied material, so every installation command cannot be independently verified here.
- The source provides no test results or complete rollback procedure for this individual skill.
How to install
This skill is part of Microsoft's Agent Skills collection. Install the collection with npx skills add microsoft/skills and select the required skills in the wizard. Alternatively, in Copilot CLI, add the marketplace with /plugin marketplace add microsoft/skills, then install the azure-skills plugin with /plugin install azure-skills@skills. The supplied source does not document a command for selecting only airunway-aks-setup.
How to use
Start with an existing AKS cluster and use a trigger such as “setup AI Runway on my AKS cluster,” “deploy a model to AKS,” or “airunway setup.” The default flow runs steps 1 through 6 in order; provide skip-to-step N to resume from a later phase. Confirm before any installation or deployment action. If no AKS cluster exists, hand off first to azure-kubernetes.
Compared to similar skills
It complements the related azure-kubernetes skill: this skill configures AI Runway on an existing AKS cluster, while azure-kubernetes handles cluster setup when one is missing. azure-diagnostics is the named companion for troubleshooting.