Modelplane v0.5: An AI gateway, fleet telemetry, and Civo

Modelplane v0.5 is out. The gateway on your control plane is now an AI gateway: it authenticates callers, reads the model a request asks for, and fails over between the backends that serve it. Modelplane also collects your fleet's metrics under one set of names, whatever engine produced them. And Civo joins the clouds Modelplane can provision a cluster on.

A ModelDeployment can also scale to zero replicas now, so one you aren't serving from costs you no GPUs. Here's what's new.

The fleet gateway is an AI gateway#

The gateway on the control plane used to be an HTTP router. It understood nothing about the requests it forwarded. A caller reached a ModelService by path prefix, nothing authenticated them, and the hop out to each cluster crossed the public internet in plain HTTP.

It's now an Envoy AI Gateway, the same one every InferenceCluster already runs at its edge. It reads the model a request names in its body and resolves the ModelService that serves it. For the backend it picks, it rewrites the model name, the credential and the path, so a backend sees the name it knows and a caller's key never reaches a third party.

ModelService gains priority alongside weight. Weight splits traffic between backends at one priority; priority fails over to the next when they go unhealthy. Every request meters a token count per caller, streams included.

InferenceGateway also stops being a singleton. It names the cluster it runs on, so you can run one per region for residency, or two in a region for availability:

apiVersion: modelplane.ai/v1alpha1
kind: InferenceGateway
metadata:
  name: eu
spec:
  clusterName: gw-gcp-eu
  tls:
    certificateRefs:
    - name: eu-example-com-tls
  auth:
    method: APIKey
    apiKey:
      secretSelector:
        matchLabels:
          modelplane.ai/inference-keys: "true"
  serviceSelector:
    matchLabels:
      example.org/region: eu

The hop from a fleet gateway to a cluster gateway is now authenticated in both directions by a per-cluster PKI, which cert-manager issues and trust-manager distributes.

One vocabulary for a fleet's metrics#

Modelplane doesn't own your engine. You bring the image and the command, and that's the point: a ModelDeployment runs vLLM, SGLang, or anything else that speaks the OpenAI API, without Modelplane knowing anything about it.

That same freedom is what makes a fleet hard to watch. vLLM publishes vllm:num_requests_waiting. SGLang calls the same measurement sglang:num_queue_reqs. DCGM reports framebuffer memory in mebibytes under a name that says bytes, and energy in millijoules under a name that says joules. A dashboard written against one engine is wrong on the next, and a fleet running both has no fleet-wide number at all. We couldn't fix that by picking an engine, so we fixed it at collection.

Modelplane now runs an OpenTelemetry collector on every inference cluster. It discovers every component Modelplane installs, renames each one's series into a single modelplane_* vocabulary, and exports them wherever you say. Only modelplane_* leaves the cluster: a series nobody renamed is one whose meaning Modelplane can't vouch for across engines, and it costs the same to carry as one that was.

Two new kinds. A TelemetryDestination says where metrics go, and nothing is collected until one exists:

apiVersion: modelplane.ai/v1alpha1
kind: TelemetryDestination
metadata:
  name: default
spec:
  sinks:
  - name: prometheus
    type: prometheus_remote_write
    endpoint: https://prom.example.internal/api/v1/write

A MetricMapping says what a component emits and what Modelplane calls it. Modelplane ships mappings for vLLM, SGLang, the gateway, the endpoint picker and DCGM, so those need nothing from you. Write one for an engine Modelplane has never seen and its numbers join the same surface:

apiVersion: modelplane.ai/v1alpha1
kind: MetricMapping
metadata:
  name: my-engine
spec:
  metrics:
  - from: my_engine_queued_requests
    to: modelplane_requests_waiting
    acrossReplicas: Sum
  - from: my_engine_kv_transfer_ms
    to: modelplane_request_kv_transfer_seconds
    fromUnit: Milliseconds
    acrossReplicas: Mean

Two fields there are worth explaining, because both encode something a dashboard would otherwise have to guess.

fromUnit exists because a metric's name is no guide to its unit. Modelplane converts to the base unit the target name claims, histogram buckets and all. Skipping it is the expensive mistake: a series named _seconds holding milliseconds reads a thousand times fast, and nothing downstream can tell.

acrossReplicas exists because every replica publishes its own series, and a query over a deployment has to combine them. Whether that's a sum or an average is a property of the measurement rather than of the query — summing two replicas at half their KV cache reads as one at full. The mapping says which, so the dashboard doesn't have to decide. Modelplane deliberately doesn't combine them in the collector: a scrape of one replica is one batch, and adding readings taken at different moments is not the traffic that happened. Your backend holds every replica's series and combines them at query time, where the arithmetic is right.

Civo#

Civo joins EKS, AKS, GKE, Nebius and Vultr as a cloud Modelplane can provision an InferenceCluster on, with the same spec you'd write for any of them.

Two things about Civo needed handling underneath. Its GPU images carry no NVIDIA driver, so the serving stack installs the GPU Operator to supply one, with the toolkit and device plugin switched off so the DRA driver stays the only thing allocating GPUs. And Civo has no server-side autoscaler, so a pool with a maxNodeCount is scaled by the upstream cluster-autoscaler running on the cluster itself. Civo's volumes are ReadWriteOnce, so ModelCache isn't available there yet.

Scaling to zero#

A ModelDeployment can now scale to zero replicas. spec.replicas used to carry a floor of one, so kubectl scale --replicas=0 was rejected at admission — which is awkward, given that scaling to zero is most of the reason to put KEDA in front of a GPU workload in the first place. There was no way to park a deployment either: withdrawing its endpoints while keeping the object meant tainting the cluster hosting it.

Dropping the floor on its own would have made a parked deployment look broken. Zero desired replicas against an empty schedule reads as none of them scheduled, and on a control plane with no clusters it reads as having nowhere to run, so a deployment you had deliberately parked would sit there permanently not ready.

Zero now takes a path of its own. Nothing is composed, so the deployment's ModelReplicas and ModelEndpoints are pruned, status.replicas reports 0 to the scale subresource, and readiness reports true with a ScaledToZero reason — the way a Deployment at zero replicas still reports Available. A parked deployment reads as parked rather than as failing, which is what makes it safe for an autoscaler to do on your behalf.

Try it#

The getting-started guide covers standing up a fleet, and Monitor the Fleet covers pointing telemetry at a backend you already run. Modelplane is Apache 2.0 and moving fast at github.com/modelplaneai/modelplane, and questions are welcome in Slack.