Overview

The ai-inference module (phase 2.1) provides a minimal inference ordering path in Deckhouse Kubernetes Platform:

1. A cluster administrator creates InferenceServiceClass — exposure policies, a required non-empty modelPolicy.allowedEndpointTypes, and optionally an acceleratorPolicy bounding the devices its orders may use. The module ships bundled default-llm.
2. A namespace user creates InferenceService with a class reference and a model, and nothing else: the platform picks the API contract from the allow-list of the class and the runtime from the recipe of the launch plan.
3. The controller validates the order, deploys the order workload objects itself, and publishes status.endpoint.

Public API: group ai.deckhouse.io/v1alpha1, kinds InferenceServiceClass (cluster) and InferenceService (namespaced).

Module components:

• CRDs and CR reference — crds/, docs/CR*.md;
• controller — controller image, namespace d8-ai-inference;
• Helm templates — templates/controller/**, user-authz and rbacv2.

Guides:

• User guide — order inference, read its status, find out why it waits;
• Administration guide — enable the module, author classes, watch and disable it;
• Configuration — the module settings reference;
• Custom Resources — the field reference of both kinds;
• Examples — copy-paste manifests;
• FAQ — answers by refusal reason.

Phase history — how each capability arrived: Phase walkthrough.

User-facing release notes: Release notes.

Repository engineering — code generation, demo and smoke scripts, the nightly recipe confirmation contour — lives in the repository README.md and in tools/recipe-confirm/README.md, not on this page.