GitHub
Menu

Roadmap

Release status
and what is planned.

Wingsmith is pre-release. There is no tagged release, image set, or chart yet, and APIs, CRDs, and the wire protocol can change without notice.

01 / Status

Where the project stands.

Status
Pre-releaseNo tagged release, images, or charts. Build from source into your own registry.
License
Apache 2.0The intended open-source license, included with the prepared codebase. Read it ↗
Built for
One organizationRunning its own clusters, operated by a single admin or a small team sharing that role.
Name
WingsmithThe planned public name for Hangar. The code keeps hgr, HANGAR_*, and hangar-operator for now.

02 / Built

Built today.

Serving only: every workload rendered today is a long-running model server.

  • Cluster enrollment over outbound-only connections
  • Per-device inventory with DCGM identity, health, and topology
  • Placement of each model on one or more clusters
  • NVIDIA Dynamo deployments with vLLM, SGLang, or TensorRT-LLM
  • An OpenAI-compatible gateway with weighted or ordered routing
  • Endpoints, tenants, API keys, and project quotas
  • Scale, pause, and drain as tracked operations
  • Discovery and adoption of models already running
  • A web console and the hgr CLI on one HTTP API
  • Conformance, benchmarking, and request-path tools

03 / Limitations

Known limitations.

Behavior that is true of the current code and worth knowing before you run it.

  • Editing a disaggregated workload's roles or device shape reloads every worker.Dynamo hashes one worker spec across the graph, so both roles restart and reload weights, which can take tens of minutes for a large model. The old generation stays up during the rollout, so the peak device need is both footprints together. The API returns this cost with the edit.
  • The gateway cannot start while the plane is down.Its routing table is held in memory only. A running gateway keeps serving; a restarted one waits for the plane.
  • The gateway is one failure domain.It is a single deployment, so if every replica is down every model is unreachable through it. Its internal listener has no authentication and must stay off any ingress.
  • The metrics cluster label is not authenticated.Any cluster with write credentials can write series under another cluster's label.
  • One admin role.Human access is a single operator with full rights. There are no per-user accounts, roles, or SSO.
  • One serving runtime.Clusters must run NVIDIA Dynamo, and workloads are rendered only as Dynamo graphs.
  • The CRD API group is interim.CRDs are v1alpha1 with no conversion, and the group name may change before the first release.

04 / Planned

Planned, not built.

Each item is reserved in the object model or designed in outline, and none of it works today. There are no dates; order within a group is roughly the order the work is expected in.

Workloads
  • Task and gang workloadsBatch work that runs to completion, and N devices together or none for training and multi-node jobs. Reserved in the object model, refused at creation today.
  • PreemptionPriority classes carry a preemptible flag, but nothing evicts anything yet.
  • Quota floors and fair shareQuotas are ceilings today. Guaranteed floors and sharing idle capacity by weight are not built.
  • Engines other than NVIDIA DynamoThe wire carries domain objects so a second renderer is possible. Dynamo is the only one.
  • Endpoint-level benchmarksBenchmarks run at the deployment and workload levels; nothing yet drives a run through the gateway.
Serving
  • AutoscalingWorkloads declare min, max, and desired scale; nothing reads the maximum yet.
  • Automatic multi-cluster failoverRouting splits by weight or sends traffic to the first live upstream. Failure detection from errors or latency and automatic cutover are not built.
  • Serving quota enforcementRequest rate, token rate, and concurrency per endpoint can be declared and shown, but the gateway does not enforce them yet.
  • Regional gatewaysThe gateway is one deployment, so one failure domain for inference everywhere.
  • Gateway cold start without the planePersisting the last known routing table would let a restarted gateway start during a plane outage.
Security and access
  • Roles, scoped tokens, and SSOHuman access is one admin role today.
  • Per-purpose cluster credentialsOne enrollment token covers the operator, the agent, and by default the metrics agent.
  • Enrollment token rotationNo rotation flow with an overlap window yet; replacing a token means minting a new one and upgrading the chart.
  • Authenticated metrics labelsA label-enforcing proxy so a cluster can only write series under its own label.
Packaging and operations
  • A Helm chart for the control planeThe cluster-setup chart covers the cluster side only.
  • Published images and chartsNothing is published yet. Build the images into your own registry.
  • Stable CRD versionsCRDs are v1alpha1 with no conversion webhooks; upgrades may require recreating them.
  • Metrics from the hub and reactorOnly the API process serves /metrics today.

Status

Open source,
under Apache 2.0.

Wingsmith is pre-release. The first public release, and everything after it, will be under Apache 2.0.

View on GitHubRead the docs Run it locally

Built and not built

Built

Multi-cluster placement and routing, Dynamo deployments, device inventory, endpoints and quotas, the console, the CLI, and three measurement tools.

Not built

Published images and charts, per-user roles and SSO, batch and training workloads, and a Helm chart for the control plane.

Pre-release. No tagged release, images, or charts yet, and APIs and protocols can change.