Run

Runhouse is the developer of Kubetorch, an open-source framework that lets engineers programmatically build, iterate on, and deploy machine learning applications on Kubernetes directly from Python. Kubetorch exposes a cluster's compute to a local Python environment, so functions defined locally can be dispatched to freshly launched remote compute on a Kubernetes cluster; logs, exceptions, and hardware faults are propagated back to the user in real time.

Because Kubetorch has no local runtime and does not rely on code serialization, it can be used from IDEs, notebooks, CI pipelines, or production code in a manner similar to a local process pool. The project describes iteration cycles of 1-3 seconds for complex workloads such as reinforcement learning and distributed training, compute cost reductions through resource allocation, bin-packing and dynamic scaling, and reduced production faults through built-in fault handling with programmatic error recovery and resource adjustment. It is distributed as a Python client package (pip install "kubetorch[client]") plus a Kubernetes-side deployment installed via a Helm chart published to an OCI registry, and is released under the Apache 2.0 license.

Business model

Kubetorch is distributed as open-source software under the Apache 2.0 license, with a fully managed serverless platform ("Kubetorch Serverless") offered by the company to interested users on request.

Traction

The public kubetorch GitHub repository has roughly 1,200 stars, 60 forks, and more than 2,500 commits, and the Helm chart is published at version 0.5.0.

Latest developments

The kubetorch repository has been consolidated to include customer-facing open-source deployment components (SDK, Helm chart, controller and data store services, workload base images, and release tooling) that were previously split between internal and open-source repositories.