related, has anyone had success moving from Slurm to Kubernetes for a physical (non-cloud) cluster primarily used for training large models on lots of GPUs?