Curious on your take of designing your own api rather than using Kubernetes. Is it due to difficulties of containerizing gpus or is the api space of k8s too bi

Curious on your take of designing your own api rather than using Kubernetes. Is it due to difficulties of containerizing gpus or is the api space of k8s too big and difficult for a small cloud provider to implement? Maybe the workloads completely don’t make sense together. I’m just curious what others think about it.