Specify GPU resource limits in the Kubernetes executor
Overview
Users can specify GPU resources when using the runner on any cloud provider-hosted Kubernetes service that supports running containers on GPU-enabled nodes.
Proposal
-
The proposed implementation plan is dependent on completion of the Spike to implement Pod Spec primitives. In the current spike MR, we are exploring the possibility of passing a PodSpec in either JSON or YAML inside the GitLab Runner config.toml. This approach enables users to modify the final PodObject sent to the Kubernetes master for scheduling.
-
A second option is to add support for Kubernetes
extended resources. A similar pattern was proposed in the enable GPU limit requests for Kubernetes executor mr.
Links to related issues and merge requests / references
Edited by Darren Eastman