From the source
Krea describes how it uses dynamic taints driven by a PromQL signal, the Descheduler, and Kueue to keep GPU workloads in the cluster and spill only when capacity is exhausted, without modifying the default Kubernetes scheduler.
From the source
From the source

From the source
So we built our design around a dynamic mechanism: detect GPU usage in the cluster, and add or remove a taint on the VK node accordingly.
krea.ai