Cozystack Passes CNCF Kubernetes AI Conformance

Cozystack passes Kubernetes AI Conformance — all twelve CNCF requirements met

Cozystack meets all twelve requirements of the CNCF Kubernetes AI Conformance programme. The v1.35 self-assessment, filed by Ænix for Cozystack v1.6.1, is accepted and published in the CNCF repository at v1.35/cozystack.

Base Kubernetes conformance answers “is this real Kubernetes”. AI conformance answers a more practical question: will an AI workload that runs on one conformant platform run here too, without platform-specific workarounds.

The requirements span accelerators, networking, scheduling, observability, security and operators. On a tenant Kubernetes cluster, Cozystack covers them with:

  • the Dynamic Resource Allocation API and the NVIDIA GPU Operator as a cluster addon;
  • GPU sharing through MIG partitions or HAMi time-slicing;
  • GPUs attached to virtual worker nodes, declared in the node pool definition;
  • Gateway API with weighted and header-based routing for model serving;
  • gang scheduling with Kueue and node pools that scale on GPU demand, down to zero;
  • accelerator and workload metrics collected through DCGM and VMAgent.

Because tenant workers are separate virtual machines with their own kernel, a GPU attached to one tenant’s node pool is not reachable from another tenant’s workloads.

Each requirement, the mechanism behind it and the command that verifies it are on the AI Conformance page.

Join the community