Your default Kubernetes scheduler grabs 4 of 16 GPUs for a distributed training job, then waits. Those 4 GPUs sit idle, reserved by pods that can’t progress unt

Your default Kubernetes scheduler grabs 4 of 16 GPUs for a distributed training job, then waits. Those 4 GPUs sit idle, reserved by pods that can’t progress until all 16 are ready, and unavailable to anyone else on the cluster. Classic partial-scheduling deadlock. kube-scheduler doesn’t gang-schedu…