Wingsmith is an open control plane for serving models across several Kubernetes GPU clusters. It enrolls your clusters, keeps a per-device inventory of them, places each model on one or more clusters, renders it there as an NVIDIA Dynamo deployment with vLLM, SGLang or TensorRT-LLM as the engine, and routes OpenAI-compatible requests to it through its own gateway.
- Two planes. The control plane stores intent. Inference traffic never passes through it, and a running gateway keeps serving from its cached routing table if the plane goes away.
- Outbound only. Each cluster's operator and agent dial out. The plane holds no kubeconfigs and nothing dials into a cluster.
- Per-device inventory. DCGM identity and health, with NVLink, NVSwitch and PCIe topology.
- Operations with a lifecycle. Scale, pause and drain have progress, an actor, a reason and a failure mode.
- Measurement built in. API conformance against a reference provider, four load generators side by side, and request-path measurement without a GPU.
Status: pre-release. No tagged release, images or charts yet, and APIs and protocols can change. Licensed under Apache 2.0.
Website · Docs · Architecture · Quickstart · Model directory · Roadmap
Built by Aircloud.