Add a workload

Add a workload when an existing integration cannot execute the task you need. Start with a nearby implementation and preserve CloudAI’s common execution behavior.

Choose the smallest extension

For an existing script or container, first consider Bash Command or Slurm Container. For a reusable integration, define its configuration and the execution strategy for each supported system.

Implement and register

1. Define command arguments and a test definition using the public interfaces in cloudai.core.

2. Implement command generation for the intended system. Reuse base strategies for allocations, mounts, and hooks.

3. Declare required installables and expose metrics from workload output where applicable.

4. Register the definition and strategies in src/cloudai/registration.py using Registry.

5. Add a public scenario, focused tests, and a workload guide. Update the compatibility catalog.

Use src/cloudai/workloads/fio for an integration with Standalone and Slurm execution, or src/cloudai/workloads/nccl_test for collective communication. Inspect the current registry signatures before adding registrations; a workload name alone does not register support for every scheduler.

Design for adaptation

Pass workload options through where practical. Keep scheduler orchestration in the execution strategy and metric interpretation in the reporter. Use mounts, installables, and hooks for environment preparation rather than hard-coding cluster details. Do not change the meaning of existing user configurations.

Verify

Mirror new Python modules under tests. Validate a minimal scenario, inspect generated commands, and cover distinct execution shapes and failure behavior. Test only the schedulers declared as supported. Follow Develop and contribute and Add a reporter for contribution and metric integration details.