This spike aims to explore the interaction between Podman's kube play functionality and the mounting of GPUs configured with CDI. Modern GPU integration often relies on CDI specifications to expose device capabilities to containers. The primary goal is to understand if Podman, when orchestrating pods from Kubernetes YAMLs via kube play, can correctly interpret and apply CDI configurations to allow GPU devices to be accessible and usable within the containers inside those pods.

The investigation should cover:

• CDI Support: Does Podman inherently support CDI for device exposure, particularly in the context of kube play?

• NVIDIA GPU Integration: Specific focus on NVIDIA GPUs, as they are a common use case for CDI. This includes checking for necessary drivers, runtimes (e.g., nvidia-container-runtime), and Podman configurations.

• kube play Compatibility: How does kube play handle Kubernetes YAMLs that specify GPU resources or CDI devices? Are there specific annotations or fields required?

• Troubleshooting: Identify common pitfalls, error messages, and potential workarounds or missing features.

• Documentation: Gather information and outline steps for successful configuration or identify limitations.

Deliverables

• Research Summary: A brief document outlining findings on Podman's CDI support, particularly with kube play.

• Proof-of-Concept (if feasible): A working example (e.g., a simple container that can detect an NVIDIA GPU) demonstrating successful GPU mounting via kube play, or clear steps outlining why it's not currently possible.

• Identified Gaps/Limitations: A list of any features or configurations missing in Podman or the environment that prevent seamless GPU integration via kube play.

• Recommendations: Suggestions for future development, necessary documentation updates, or alternative approaches.