Skip to content

feat: add NIM Operator recipe for CNCF AI Conformance certification #473

Description

@yuanchen8911

Summary

Based on CNCF AI Conformance WG feedback, the conformance submission should certify a commercial product on a specific K8s distribution, not standalone tooling. We will convert the existing AICR conformance PR to certify NIM on EKS.

Background

Discussion in k8s-ai-conformance Slack:

  • CNCF conformance requires submitting a product per K8s offering (e.g., NIM on EKS, NIM on AKS)
  • AICR serves as the validation tooling (analogous to sonobuoy for K8s conformance), not the certified product
  • The certified product is NVIDIA NIM deployed on EKS

Proposal

Add a new AICR recipe that deploys NIM on EKS using the k8s-nim-operator:

  1. New components in recipes/registry.yaml:

    • k8s-nim-operator — Helm chart from helm.ngc.nvidia.com/nvidia/helm-charts/k8s-nim-operator
    • Deploys the NIM Operator which manages NIM lifecycle (autoscaling, health monitoring, model caching)
  2. New recipe overlay: h100-eks-ubuntu-inference-nim.yaml

    • Inherits from h100-eks-ubuntu-inference
    • Adds k8s-nim-operator component
    • Includes a NIMService CR to deploy a small model (e.g., Llama 3.2 1B) for conformance validation
    • Full conformance validation suite (all 11 checks)
  3. Conformance PR conversion:

    • Update the existing CNCF conformance PR to use the NIM recipe
    • Evidence collected against NIM on EKS deployment
    • Follow-up submissions for GKE, AKS, OKE, etc.

Why NIM Operator (not raw NIM Helm chart)

  • Production deployment method with full lifecycle management
  • Natively supports autoscaling and health checks required by conformance
  • NIMService CR demonstrates the product experience, not just a container
  • Stronger certification narrative: "NIM Operator on EKS"

Tasks

  • Add k8s-nim-operator component to recipes/registry.yaml
  • Create component values at recipes/components/k8s-nim-operator/values.yaml
  • Create NIMService CR manifest for conformance test workload
  • Create h100-eks-ubuntu-inference-nim.yaml recipe overlay
  • Validate conformance checks pass against NIM deployment on EKS
  • Convert CNCF conformance PR to use NIM on EKS

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions