Kubetorch

Kubetorch is a modern interface for running heterogeneous ML workloads on Kubernetes. While Kubernetes is the standard compute foundation, development and deployment remain challenging. Notebook-style or VM-based workflows slow productionalization, limit distributed scaling, and make inference difficult. Direct development on Kubernetes requires complex YAML and long iteration cycles, even for small changes.

With Kubetorch, you write programs in regular Python and can develop and debug interactively while running at scale on Kubernetes. Kubetorch's magic caching and deployment system enables nearly instanteous relaunch of your programs. Meanwhile, your code executes consistently across any environment, whether a teammate's laptop, CI, an orchestrator, or a production application.

Kubetorch combines a Python library with a Kubernetes operator deployed in your cloud account to provide a simple, flexible, and powerful way to build, deploy, and manage AI/ML applications. It can be adopted incrementally within an existing ML stack or used as a full replacement for training, batch processing, inference, hyperparameter optimization, and pipelining tools such as Kubeflow, SageMaker, or Vertex.