Kubernetes GPU Scheduling for MLOps and GPU Sharing

Introduction The default Kubernetes scheduler was built for stateless services: web servers, APIs, databases. It schedules a pod, checks that a node has enough of whatever resources were requested, and binds it. For CPU and memory, that model works fine.