Modelplane 开源：在自有环境编排模型与推理基础设施

Modelplane is software you install and run in your own environment to orchestrate models, the serving stack, and the infrastructure underneath across cloud, neocloud, and on-premise. It runs any model on any engine on any infrastructure, from a single GPU to disaggregated, multi-node deployments. Built on Crossplane, it is an active system that continuously reconciles your fleet toward the state you declare: provisioning inference clusters, scheduling deployments onto compatible clusters, scaling replicas, caching weights, and routing traffic.

Platform teams provision clusters and publish hardware classes. Developers declare a model and get back a unified, OpenAI-compatible endpoint. Neither team has to know the details of the other's job.

Warning

Modelplane is an early release under active development. Its APIs and behavior can change between releases. We are building it in the open, collaborating with the AI inference community on integrations and capabilities.

Deploy a model

Once a platform team has provisioned inference clusters and declared the available GPUs, a developer deploys a model with a declarative manifest:

apiVersion: modelplane.ai/v1alpha1
kind: ModelDeployment
metadata:
  name: qwen-demo
  namespace: ml-team
spec:
  replicas: 1
  template:
    spec:
      engines:
      - name: qwen
        members:
        - role: Standalone
          nodeSelector:
            devices:
            - name: gpu
              count: 1
              selectors:
              - cel: device.capacity["gpu.nvidia.com"].memory.compareTo(quantity("20Gi")) >= 0
          template:
            spec:
              containers:
              - name: engine
                image: vllm/vllm-openai:v0.23.0
                args:
                - --model=Qwen/Qwen2.5-0.5B-Instruct
                - --served-model-name=$(MODELPLANE_SERVED_MODEL_NAME)

Modelplane schedules the replica onto a cluster with free, compatible GPUs and deploys the serving engine. Give callers one model name for it at the gateway with a ModelService:

apiVersion: modelplane.ai/v1alpha1
kind: ModelService
metadata:
  name: qwen
  namespace: ml-team
spec:
  endpoints:
  - name: qwen-demo
    selector:
      matchLabels:
        modelplane.ai/deployment: qwen-demo

Getting started

Follow the getting started guide to deploy Modelplane on a local kind cluster and serve a model. The how it works page covers the resources and what happens when you deploy a model.

The recipes are validated, end-to-end manifest sets that serve specific models, each covering the full workflow from inference class and cluster through model cache, deployment, and service.

How it works

Modelplane runs as a control plane on its own cluster, above the inference clusters that serve models. Its API is two sets of resources, one per role, with everything in between composed for you:

• Platform teams create InferenceClusters (the GPU fleet, provisioned by Modelplane or brought as-is) and InferenceClasses (hardware recipes: the devices a node pool offers and how to provision it), fronted by an InferenceGateway.
• Developers create a ModelDeployment (a model's engines, replica count, and an optional ModelCache) and a ModelService (one model name, at the gateway, across the replicas it selects).
• Modelplane composes a ModelReplica per cluster and a ModelEndpoint per replica.

Once those resources exist, Modelplane keeps the fleet matching them across five concerns: provisioning clusters and their node pools, scheduling each replica onto a cluster and pool whose hardware fits, scaling replicas through the standard Kubernetes scale subresource, routing through one OpenAI-compatible endpoint with weighted canary and A/B rollouts, and caching model weights once per cluster.

Modelplane is unopinionated about the engine. A ModelDeployment describes the shape of a deployment, how many pods, on how many nodes, with which devices; the engine flags you write carry parallelism, quantization, and KV transfer. Modelplane never injects them. That is what lets one API serve any container-based engine and any topology.

Current status

Modelplane is at v0.1. It is early and evolving fast. See issues labeled enhancement for what's planned.

Get involved

Contributions, bug reports, and feature requests are welcome.

• Issues: GitHub Issues
• Slack: Join our Slack

See CONTRIBUTING.md for how to get set up, run checks, and submit changes.

License

Modelplane is under the Apache 2.0 license.

Modelplane™ is a trademark. The Apache 2.0 license grants no trademark rights: the Modelplane name and logos are not covered by it.