rhoai-hardware-profiles

name	rhoai-hardware-profiles
metadata	{"author":"rhoai3-demo","version":"1.0.0","platform-family":"rhoai","platform-baseline":"repo","ocp-baseline":"repo","skill-group":"RHOAI Platform"}
description	Use when documenting, reviewing, or rebuilding Red Hat OpenShift AI hardware profiles from the official Working with accelerators guide: creating, updating, deleting, enabling, disabling, or GitOps-promoting HardwareProfile resources; setting accelerator identifiers, resource requests and limits, node selectors, tolerations, and Kueue local queues; reviewing recommended accelerator tags for workbench images and serving runtimes; validating dashboard visibility for workbenches and model serving; and aligning NVIDIA L40S demo profiles with active node capacity. Do NOT use for GPU node provisioning, NFD, NVIDIA GPU Operator, ClusterPolicy, or GPU time slicing setup; use rhoai-nvidia-gpu-accelerators and related ocp-* skills. Do NOT invent HardwareProfile spec fields; verify active schema or export a dashboard-created profile before authoring GitOps manifests.

RHOAI Hardware Profiles

Use this skill to create, review, and promote OpenShift AI hardware profiles on the active product baseline in docs/PLATFORM_BASELINE.md.

Source Grounding

Read references/source-capture.md before using product configuration details. Official Red Hat documentation is product authority. The RHOAI 3.4 accelerator guide defines the dashboard workflows and validation signals for hardware profiles. Use rhoai-api-tiers for API support posture before treating hardwareprofiles.infrastructure.opendatahub.io/v1 as a durable contract.

Scope

This skill covers:

• hardware profile creation, update, deletion, enablement, and disablement
• accelerator identifiers such as nvidia.com/gpu
• CPU, memory, and accelerator resource request and limit intent
• Kueue local queue association when distributed workload scheduling is used
• workload priority selection
• node selector and toleration intent
• recommended accelerator tags for workbench images
• recommended accelerator annotations on serving runtimes
• GitOps promotion of verified HardwareProfile resources

This skill does not cover:

• installing NFD, KMM, NVIDIA GPU Operator, or accelerator Operators
• AWS GPU MachineSets or GPU node lifecycle
• NVIDIA ClusterPolicy, GPU time slicing, or RDMA configuration
• full workbench image or serving runtime authoring
• queue design beyond the hardware profile handoff

Demo Policy

For this repo:

• Default accelerator identifier: nvidia.com/gpu.
• Default demo hardware shape: one NVIDIA L40S GPU per g6e.2xlarge GPU worker, two workers, advertised as one nvidia.com/gpu unit each — this repo disables time-slicing so every private model claims an exclusive card (see stages/020 README for the rationale).
• Default profile intent: queue-backed GPU profiles for workbenches and the Nemotron + Qwen3.6 model-serving workloads unless a stage README documents a different resource class.
• Create hardware profiles only after GPU support is installed and nodes report accelerator capacity and allocatable resources.
• Use active node labels and taints as scheduling authority. Do not infer selectors or tolerations from AWS instance type alone.
• If Kueue is used, coordinate with rhoai-kueue-workload-management; do not combine Kueue local queue scheduling with conflicting node selectors or tolerations unless verified against the active docs and CRD schema.

Workflow

1. Confirm the active baseline in docs/PLATFORM_BASELINE.md.

2. Confirm accelerator prerequisites are complete with rhoai-nvidia-gpu-accelerators, ocp-node-feature-discovery, ocp-machine-management, and ocp-nodes as needed.

3. Read references/official-doc-extraction.md.

4. Decide whether the profile is global or project-scoped:

   • global profiles belong with the shared RHOAI platform owner
   • project-scoped profiles belong with the consuming project stage and must be coordinated with rhoai-project-scoped-resources

5. For dashboard-first creation, create the profile from Settings -> Hardware profiles, then export the resulting HardwareProfile before promoting to GitOps.

6. For GitOps-first authoring, verify schema first:

   oc explain hardwareprofile --api-version=infrastructure.opendatahub.io/v1
   oc explain hardwareprofile.spec --api-version=infrastructure.opendatahub.io/v1
   oc get crd hardwareprofiles.infrastructure.opendatahub.io -o yaml
   

7. Capture resource intent: accelerator identifier, CPU/memory defaults, accelerator defaults, limits, local queue, workload priority, node selectors, and tolerations.

8. Add recommended accelerator metadata to workbench images or serving runtimes only when it matches the active profile and runtime/image compatibility.

9. Validate with references/validation-checklist.md.

Validation

• The hardware profile appears on the dashboard Hardware profiles page.
• The profile appears in the Create workbench hardware profile list when it is intended for workbenches.
• The profile appears on the HardwareProfile CRD instances page.
• GPU-backed profiles match observed node capacity and allocatable resources.
• Recommended accelerator tags appear only where the accelerator is actually compatible with the workbench image or serving runtime.
• GitOps manifests render and match the active CRD schema.

Related Skills

• rhoai-nvidia-gpu-accelerators: accelerator installation, NVIDIA-only demo policy, GPU MachineSet handoff, and GPU Operator readiness.
• rhoai-project-scoped-resources: project-scoped hardware profile copies and dashboard disableProjectScoped behavior.
• rhoai-kueue-workload-management: local queue scheduling and Kueue-enabled hardware profile restrictions.
• rhoai-model-deployment: hardware profile selection during model deployment.
• rhoai-workbench-image-import: workbench image compatibility and accelerator metadata.
• rhoai-model-serving-platform: serving runtime compatibility and Dashboard behavior.

References

• references/source-capture.md
• references/official-doc-extraction.md
• references/validation-checklist.md
• examples/nvidia-l40s-hardware-profile-pattern.md