kubetorch

Why we included this project

Anyone who has run real ML on a cluster knows the wall: you iterate in Python, then hit a stack of Kubernetes manifests and scheduling delays just to test a training or evaluation job. Kubetorch hides that plumbing behind a plain Python API, letting your cluster act like a local process pool you can call from notebooks, IDEs, CI pipelines, or production code. Write a normal function, attach a Compute spec, and the tool launches remote workers, streams logs, exceptions, and hardware faults back in real time, then reclaims resources when the job finishes. That makes it practical to iterate on distributed training, RL, or other compute-heavy workloads without babysitting pods. Underneath, it does resource bin-packing and dynamic scaling, which covers much of the cost and reliability work, so you get cluster-scale compute without becoming a Kubernetes expert.