Wano

Local-first compute orchestration for multi-node CPU/GPU workloads. Request abstract compute types, not specific devices. Automatic node discovery, distributed execution via Ray, self-hosted with no cloud dependency.

graph TD
    User[User] -->|Submit Job| API[Control Plane<br/>FastAPI + Scheduler]

    subgraph "Control Plane"
        API <--> DB[(SQLite DB)]
        RayHead[Ray Head]
    end

    Worker[Worker Nodes] -->|Discover via mDNS & Register| API
    Worker -->|Join Cluster| RayHead

    API -->|Schedule & Execute| RayHead
    RayHead -->|Run Tasks| Worker

Loading

CLI Commands

• wano up - Start the control plane (runs detached by default)
• wano down - Stop the control plane
• wano join - Register this machine as a worker node
• wano status - View cluster status and active jobs
• wano run <script> --compute <cpu|gpu> [--gpus N] [--function NAME] [--args JSON] [--kwargs JSON] - Submit and execute a job
• wano build-executor [--requirements requirements.txt] - Build the executor Docker image

Python API

import wano

@wano.function(compute="gpu", gpus=4)
def train():
    # Runs on 4 GPUs, potentially across multiple nodes
    ...

@wano.function(compute="cpu")
def preprocess():
    # Runs on CPU resources
    ...

Executor Image

The executor image must include any Python dependencies used by your jobs. Build it once per project and reuse it:

wano build-executor --requirements requirements.txt