cncf/k8s-ai-conformance

What it solves

It addresses the lack of standardization across Kubernetes platforms when running AI and machine learning workloads. Because AI workloads have unique requirements for accelerators, networking, and scheduling, they often suffer from "it works on my cluster" issues. This project provides a certification program to ensure AI applications are portable across different conformant platforms.

How it works

The program defines a set of capabilities that a Kubernetes platform must possess to be considered AI-conformant. Vendors undergo a certification process that includes:

1. Prerequisite: The platform must first be base Kubernetes Conformant.
2. Verification: Vendors complete a conformance checklist and provide evidence (documentation or URLs). For newer versions (v1.37+), they are encouraged to run an automated test suite to generate artifacts like e2e.log and results.json.
3. Submission: Vendors submit a pull request containing their checklist and evidence to the CNCF for review.
4. Certification: Once approved, the platform is certified for one year.

Who it’s for

• Platform Vendors: Those providing Kubernetes-based platforms who want to certify their infrastructure for AI/ML workloads.
• Contributors: Developers and researchers who want to help define the requirements for AI workload types, including training, inference, and agentic workflows.
• AI Application Developers: Users who want a reliable baseline to build and test against to ensure their apps work across different clusters.

Highlights

• Broad Workload Coverage: Focuses on distributed training, LLM inference, and multi-step agentic workflows.
• Hybrid Verification: Combines manual attestation with automated test results for faster, more objective validation.
• Vendor-Neutral Baseline: Established as a community-led effort to prevent platform-specific lock-in.