ocp-ai-workloads

name	ocp-ai-workloads
metadata	{"author":"rhoai3-demo","version":"1.0.0","platform-family":"ocp","platform-baseline":"repo","ocp-baseline":"repo","skill-group":"OpenShift Platform"}
description	Use when documenting, reviewing, or rebuilding OpenShift Container Platform AI workload platform guidance from official OCP documentation: Red Hat build of Kueue, Kueue custom resources, ClusterQueue, ResourceFlavor, LocalQueue, Workload, queue labels, quotas, RBAC, cohorts, fair sharing, gang scheduling, pending workload visibility, Leader Worker Set Operator, LeaderWorkerSet, JobSet Operator, JobSet, and coordinated AI training or inference workload scheduling. Do NOT use for RHOAI dashboard, DataScienceCluster, workbenches, or OpenShift AI distributed workload component configuration; use the relevant rhoai-* skill. Do NOT run live queue, quota, RBAC, or operator changes without the env safety guard and explicit approval.

OCP AI Workloads

Use this skill to ground OpenShift AI workload platform guidance in the official OpenShift Container Platform AI workloads guide for the active baseline in docs/PLATFORM_BASELINE.md.

Source Grounding

Read references/source-capture.md before using product behavior. Official Red Hat documentation is product authority. This skill covers OCP platform operators and Kubernetes APIs for AI workload scheduling and coordination; it does not replace RHOAI component skills, GPU enablement skills, or live environment skills.

Demo AI Workload Posture

For this AWS-hosted RHOAI demo:

• Treat Red Hat build of Kueue as an OpenShift platform scheduling and quota layer for batch, distributed, and GPU-constrained workloads.
• Use Kueue queues to make limited GPU capacity explicit before RHOAI or demo workloads consume it.
• Keep RHOAI component enablement in rhoai-kueue-workload-management, rhoai-distributed-workload-operations, and related rhoai-* skills.
• Use ocp-node-feature-discovery, ocp-nodes, and rhoai-nvidia-gpu-accelerators before claiming that GPU queues map to real GPU capacity.
• Verify installed Operator versions, channels, CRDs, API groups, and support posture before writing README claims or GitOps manifests.

OCP AI Workload Model

Use the official docs to frame:

• Red Hat build of Kueue: Kubernetes-native resource access management for jobs, including waiting, admission, and preemption.
• Kueue custom resource: cluster-level Kueue configuration, including framework integrations, preemption policy, fair sharing, and gang scheduling.
• ClusterQueue: cluster-scoped quota pool for covered resources and resource flavors.
• ResourceFlavor: named resource variation associated with node labels, taints, and tolerations.
• LocalQueue: namespaced queue that points workloads in a namespace to a ClusterQueue.
• Workload: Kueue representation of a submitted job or integrated workload.
• LeaderWorkerSet: coordinated groups of leader and worker pods for distributed inference or sharded model serving workloads.
• JobSet: coordinated groups of Kubernetes Jobs for large-scale training, HPC, and related batch workloads.

Workflow

1. Confirm the active OpenShift baseline in docs/PLATFORM_BASELINE.md.
2. Read references/official-doc-extraction.md.
3. Identify whether the task concerns:
   • Red Hat build of Kueue installation, upgrade, or disconnected install
   • Kueue custom resource configuration
   • namespace opt-in with kueue.openshift.io/managed=true
   • ClusterQueue, ResourceFlavor, LocalQueue, or Workload resources
   • RBAC with kueue-batch-admin-role or kueue-batch-user-role
   • quotas, cohorts, fair sharing, gang scheduling, or visibility API use
   • Leader Worker Set Operator or LeaderWorkerSet
   • JobSet Operator or JobSet
4. For GitOps manifests, verify API versions and CR fields against official docs or the active cluster schema before committing.
5. For live operations, use the repo environment guard and pair this skill with env-troubleshoot, env-manage-resources, or env-deploy-and-evaluate.
6. Validate the output with references/validation-checklist.md.

Related Skills

• Use rhoai-kueue-workload-management for RHOAI-managed Kueue component enablement and RHOAI-specific dashboard behavior.
• Use rhoai-distributed-workload-operations and rhoai-distributed-workload-workflows for OpenShift AI distributed workload administration and data scientist workflows.
• Use rhoai-distributed-inference-llmd for llm-d and RHOAI distributed inference.
• Use ocp-node-feature-discovery, ocp-machine-management, ocp-nodes, and rhoai-nvidia-gpu-accelerators for GPU worker capacity and placement.
• Use project-gitops-authoring and project-manifest-review for manifests.

References

• references/source-capture.md
• references/official-doc-extraction.md
• references/validation-checklist.md
• examples/ai-workload-review-patterns.md