Ray¶

Ray is for sending Wove tasks into a Ray cluster without writing a separate backend-specific task wrapper. Wove wraps its worker entrypoint as a Ray remote function and submits the payload directly.

Use When¶

Your application already uses Ray for distributed Python execution.
Wove tasks should run on Ray workers rather than the submitting process.
The Ray workers can import Wove and your application code.

Execution Shape¶

Wove initializes Ray unless init=False.
Wove wraps wove.integrations.worker.run with ray.remote(...).
The remote function executes the payload.
The worker posts Wove completion events back to the callback URL.

Dependency¶

Install dispatch support and Ray in the submitting process and on Ray workers that execute Wove payloads.

pip install "wove[dispatch]" ray

Configure Wove¶

import wove

wove.config(
    default_environment="ray",
    environments={
        "ray": {
            "executor": "ray",
            "executor_config": {
                "address": "ray://ray-head:10001",
                "remote_options": {"num_cpus": 1},
                "callback_token": "shared-secret",
                "callback_url": "https://wove-runner.internal/wove/events/shared-secret",
            },
        }
    },
)

Worker Requirements¶

Ray workers must have Wove installed and must be able to import the application code used by the serialized task.

Options¶

Key	Effect
`address`	Ray cluster address passed to `ray.init(...)`.
`init`	Set to `False` if the process already initialized Ray.
`init_options`	Extra keyword arguments passed to `ray.init(...)`.
`remote_options`	Options passed to `ray.remote(...)`.
`force_cancel`	Passed to `ray.cancel(..., force=...)`.