cncf/k8s-ai-conformance

What it solves

This project establishes a standardized approach for running AI/ML workloads on Kubernetes. It addresses the issue of platform-specific inconsistencies—where AI applications may work on one cluster but fail on another—by defining a baseline of capabilities that Kubernetes platforms must meet to be certified as "AI Conformant."

How it works

The program operates as a certification process for Kubernetes platform vendors. Vendors must first be Kubernetes Conformant and then provide a self-assessment checklist along with public evidence (documentation and test results) to prove they support the required AI/ML capabilities. Starting with v1.37, the program supports a hybrid verification approach combining automated test results (for features like Gang Scheduling and Secure Accelerator Access) with manual attestation.

Who it’s for

• Kubernetes Platform Vendors: Those seeking certification to prove their platform's reliability for AI workloads.
• AI Tooling Developers: Developers who need a clear, vendor-neutral baseline to build and test their tools against.
• End Users: Users who want to ensure their AI workloads are portable across different Kubernetes platforms.

Highlights

• Broad Workload Support: Covers training (distributed/large jobs), inference (model/LLM serving), and agentic workloads (multi-step workflows).
• Certification Lifecycle: Certifications are valid for one year and aligned with Kubernetes release cycles.
• Hybrid Verification: Combines automated tests with manual documentation review.
• Vendor-Neutral Baseline: Community-led effort to ensure AI portability across the ecosystem.