AIM model catalog management

installation AIM catalog cluster operator model catalog

AIM model catalog management#

Entry point for cluster operators and platform administrators who manage the AIM model catalog on AMD Enterprise AI reference stack clusters. End users browse and deploy models from the catalog in AMD AI Workbench.

Detailed lifecycle rules, version policy, and step-by-step Gitea and Argo CD procedures live in the Cluster Forge repository (not duplicated here).

Operator guides (Cluster Forge)#

Topic

Guide

Catalog layers, version policy, filter vs source removal, responsibilities

AIM catalog lifecycle

Add, replace, or remove entries; active hardware families; troubleshooting

AIM model management

To list active hardware families on a cluster, use the fenced kubectl command in Before you start in AIM model management (not duplicated here).

Portal configuration#

Topic

Guide

Set AIM_HARDWARE_FAMILY / hardwareFamilies at install

Installation configuration values

Set GPU_STACK_FAMILY (GPU Operator chart and DeviceConfig; separate from the catalog chart)

Installation configuration values (GPU_STACK_FAMILY)

Upgrade the reference stack; dedupe before catalog or platform upgrade

Upgrade an existing installation

At a glance#

Enterprise AI reference stack clusters combine a packaged baseline catalog (Cluster Forge / Argo CD) with optional cluster-managed additions (Gitea). Lifecycle rules, version policy, filter vs source removal, and step-by-step procedures are in the Cluster Forge operator guides above — not repeated here.

The packaged baseline is the Cluster Forge in-tree Helm chart sources/aim-cluster-model-source. The AIM team publishes model and base-image lists in aim-build; Platform copies approved snapshots into that chart at each Cluster Forge release.

At install, set AIM_HARDWARE_FAMILY to inject a YAML list into apps.aim-cluster-model-source.valuesObject.hardwareFamilies. When it is omitted, cluster-bloom leaves that list unset. A non-empty list uses templates/profiles.yaml (only listed families; the Instinct profile includes generic amd-aim-release-* 0.8.5–0.11.0 plus Instinct 0.11.1+). An empty list ([]) is the chart default and selects templates/unfiltered.yaml: Instinct 0.11.1, 0.12.0, 0.13.0 plus mixed base images. Clearing the list in Gitea after a family-filtered install switches to that unfiltered branch; it does not fail chart rendering.

Important

The aim-cluster-model-source Helm chart selects catalog sources through hardwareFamilies only. An empty hardwareFamilies list ([]) selects the unfiltered catalog. An empty apps.amd-gpu-operator-config.valuesObject.gpuStackFamily resolves to instinct (instinct or radeon). bloom.yaml GPU_STACK_FAMILY also sets the operator chart path and driverVersion. See Installation configuration values.

Before upgrading Cluster Forge, review the incoming packaged catalog and remove cluster-managed manifests that duplicate newly included models or bases. See Before a platform upgrade in the Cluster Forge how-to guide.

CR and chart references#

Topic

Guide

AIMClusterModelSource append-only discovery and source deletion

Model Sources (AIM Engine)

Packaged baseline chart (hardwareFamilies empty = Instinct 0.11.1+ plus mixed bases)

aim-cluster-model-source README (Cluster Forge main)
Contents