You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Kruize currently works along side with VPA to apply the recommendations, but VPA doesn't apply the GPU partitioning changes to apply the accelerator recommendations generated by kruize. Instaslice Operator on Openshift actually partitions the GPU on the fly when ever the pod resources for GPU are changed. So If kruize can track the MIG partition usage and come up with MIG right sizing recommendations, kruize can apply them via instaslice (As on when kruize changes the pod requests and limits wrt to GPU instaslice can create the slice and assign to the pod). This way the accelerator recommendations can also be applied via kruize.
Suggest a solution
Kruize should be able to track the MIG partition (which are created by instaslice) usage and come up with accelerator recommendations and apply them by updating the kubernetes deployment / statefulset / job object with latest requests and limits and let instaslice create the GPU slices on the fly.
Describe the feature
Kruize currently works along side with VPA to apply the recommendations, but VPA doesn't apply the GPU partitioning changes to apply the accelerator recommendations generated by kruize. Instaslice Operator on Openshift actually partitions the GPU on the fly when ever the pod resources for GPU are changed. So If kruize can track the MIG partition usage and come up with MIG right sizing recommendations, kruize can apply them via instaslice (As on when kruize changes the pod requests and limits wrt to GPU instaslice can create the slice and assign to the pod). This way the accelerator recommendations can also be applied via kruize.
Suggest a solution
Kruize should be able to track the MIG partition (which are created by instaslice) usage and come up with accelerator recommendations and apply them by updating the kubernetes deployment / statefulset / job object with latest requests and limits and let instaslice create the GPU slices on the fly.
Checklist of changes to be done
The text was updated successfully, but these errors were encountered: