Add an execution system or runner

Add a backend when CloudAI must execute experiments through another scheduler or environment.

Follow an existing backend

Read src/cloudai/systems/standalone for local process execution and src/cloudai/systems/slurm for scheduled execution. Identify the system model, runner, installer, and workload execution strategies before writing a new backend.

Connect the components

• Define the system’s configuration through the public core interfaces.

• Implement submission, polling, completion, cancellation, and result-path handling in the runner.

• Provide prerequisite installation appropriate to the environment.

• Register the system, runner, and installer with Registry in the startup registration path.

• Register workload strategies for that system and declare their compatibility explicitly.

Import shared interfaces from cloudai.core and respect the core’s dependency boundaries. An execution backend and a workload adapter solve different parts of the task; verify both together.

Verify behavior

Cover submission failure, queued and running states, completion, cancellation, and output collection. Document path visibility, resource allocation, authentication, and supported workloads. Preserve the structured experiment output consumed by Automate experiments. Follow Develop and contribute for repository conventions.