apiVersion: catalog.confighub.com/v1alpha1 kind: PlatformRouteIntent metadata: name: aicr-eks-h100-training-kubeflow-v1-0-0 spec: source: type: aicr provider: NVIDIA sourceVariant: h100-eks-ubuntu-training-kubeflow record: examples/aicr/eks-h100-training-kubeflow-v1-0-0/generation-receipt.yaml materializedObjects: examples/aicr/eks-h100-training-kubeflow-v1-0-0/argocd-rendered boundary: exactObjects: >- The retained files are 17 Argo CD Application objects. Their exact content can be reviewed and transported before delivery. downstreamRendering: >- Sixteen component Applications still reference Helm or local chart sources. Argo CD or Flux processes those sources later. This entry does not call the wrapper Applications the final workload objects. observation: >- The v1.0.0 CLI lists snapshot and diff commands for node facts. Neither was run for this entry. A reported difference is a fault only after the node is bound to the intended provider-curated variant and target role. targetFacts: - Argo CD and its Application CRD must exist before the retained wrapper objects can be applied. - The destination must match the selected EKS, H100, Ubuntu, and Kubeflow training source variant. - A gp3 StorageClass must be available for the selected storage input. - Nodes that run system components must carry the system node selector label. This bundle carries the placeholder nodeGroup=system-worker, and the destination's own label replaces it. - GPU nodes must carry nvidia.com/gpu.present=true before the node-tuning object is applied. - The GPU driver, EFA device plugin, and node tuning must suit the actual node and network role. targetFactBasis: - fact: Argo CD and its Application CRD basis: Every retained object is an argoproj.io/v1alpha1 Application. - fact: source variant match basis: The recipe criteria and its constraints (Kubernetes >= 1.32.4, Ubuntu 24.04, kernel >= 6.8). - fact: gp3 StorageClass basis: The storageClass generation input. - fact: nodeGroup=system-worker newInVersion: v1.0.0 basis: >- The systemNodeSelector generation input. It is written into the values of 12 of the 16 component Applications. A cluster without that label on any node leaves those components unschedulable. The value is a confirmed placeholder. generation-receipt.yaml records every field path it lands in, and the system-node-selector-placeholder route records how it is replaced. - fact: nvidia.com/gpu.present=true before node tuning newInVersion: v1.0.0 basis: >- The acceleratedNodeSelector generation input sets the NodeWright tuning selector to that label. The generated tuning template itself says the label comes from GPU feature discovery inside gpu-operator, which is ordered after nodewright-customizations, so on a new cluster the selector matches no node until something else applies the label. Whether that matters depends on the destination and is not assessed here. - fact: driver, EFA, and node tuning basis: >- The gpu-operator values pin driver 580.173.02 with kernelModuleType auto and driver RDMA disabled. The aws-efa component is included. The NodeWright tuning object reboots matched nodes one at a time. routes: - id: source-variant-target-match lifecycleWork: compare observed destination facts with the intended NVIDIA source variant owner: platform operator or admission policy order: before destination-specific promotion or delivery evidenceRequired: target snapshot, selected source-catalog digest, and an explained match or exception status: recorded-not-run - id: argocd-prerequisite lifecycleWork: make the Argo CD Application API and controller available owner: platform operator order: before any retained Application object evidenceRequired: Argo CD CRD discovery and controller readiness status: recorded-not-run - id: component-order lifecycleWork: reconcile component Applications in AICR dependency order owner: Argo CD order: sync-wave annotations generated from recipe.deploymentOrder evidenceRequired: observed sync operations and component health, not annotation survival alone status: recorded-not-run - id: downstream-chart-lifecycle lifecycleWork: handle CRDs, hooks, certificates, operators, and setup Jobs for every nested source owner: destination controller and the referenced source processor order: source- and destination-specific evidenceRequired: nested source render, flattening decision, route resolution, and execution receipt status: not-yet-materialized-for-v1.0.0 - id: upgrade-over-existing-installation lifecycleWork: follow the manual steps AICR lists before applying this bundle over an earlier installation owner: platform operator order: before the first sync of this version on a cluster that already runs an earlier one evidenceRequired: the completed steps for each listed component and an upgrade-check report for the exact move evidence: - examples/aicr/eks-h100-training-kubeflow-v1-0-0/argocd-helm-bundle/UPGRADING.md - examples/aicr/eks-h100-training-kubeflow-v1-0-0/flux-bundle/UPGRADING.md observed: - >- UPGRADING.md lists manual steps for nodewright-operator when moving from below 0.18.0, and says that move is not reversible. The v0.20.0 entry pins nodewright-operator v0.17.1 and this entry pins v0.19.0. - >- UPGRADING.md lists manual steps for nvsentinel when moving from 1.22.0 up to but not including 1.25.0. The v0.20.0 entry pins nvsentinel v1.20.0, which is outside that range, so the file says nothing about that exact move. - >- aicr upgrade-check from the retained v0.20.0 recipe to this bundle was refused by the v1.0.0 binary, because the v0.20.0 recipe uses apiVersion aicr.run/v1alpha2, which v1.0.0 no longer reads. No upgrade verdict from v0.20.0 exists for this entry. status: recorded-not-run newInVersion: v1.0.0 - id: system-node-selector-placeholder lifecycleWork: replace the placeholder system node selector nodeGroup=system-worker with the label on the destination cluster's system node group owner: platform operator order: before the bundle is published or delivered to a destination evidenceRequired: the label observed on the destination's system node group, and either a bundle regenerated with that value or a variant made from this entry, with its new digests evidence: - examples/aicr/eks-h100-training-kubeflow-v1-0-0/generation-receipt.yaml observed: >- The placeholder is written into spec.source.helm.values of 12 of the 17 rendered Applications, at 24 field paths that the generation receipt lists. A different value changes the bundle bytes and every digest in this entry. status: recorded-not-run newInVersion: v1.0.0 - id: dra-plugin-eviction lifecycleWork: decide whether the DRA kubelet plugin is evicted before a GPU driver container restarts owner: platform operator order: before the first GPU driver upgrade or restart evidenceRequired: a chosen DRA eviction node label and a regenerated bundle, or a recorded decision to leave it off observed: >- Bundle generation warned that no DRA eviction node label was set and left dra-node-labeler out of the bundle. The recipe still lists that component. status: recorded-not-run newInVersion: v1.0.0 - id: aicr-node-inspection lifecycleWork: collect and compare node facts owner: AICR snapshot and diff order: before or independently of configuration delivery evidenceRequired: snapshot digests and the exact diff status: available-not-run-for-this-entry - id: aicr-configuration-validation lifecycleWork: run checks whose prerequisites are satisfied by the selected source variant owner: AICR validator or the operator's test runner order: after the relevant components or hardware exist evidenceRequired: per-check prerequisite classification and target output status: not-run checkClasses: status: not-assessed-for-v1.0.0 reason: >- The v0.20.0 entry classifies AICR checks by what each needs before it can run. That classification was not re-derived against the v1.0.0 validator, so it is not repeated here. The recipe declares 4 deployment checks, 1 performance check, and 10 conformance checks. status: configurationGenerated: true routesExecuted: false runtimeProven: false