ray.train.backend.Backend

class ray.train.backend.Backend(*args, **kwargs)

Singleton for distributed communication backend.

share_cuda_visible_devices

If True, each worker process will have CUDA_VISIBLE_DEVICES set as the visible device IDs of all workers on the same node for this training instance. If False, each worker will have CUDA_VISIBLE_DEVICES set to the device IDs allocated by Ray for that worker.

Type:

bool

DeveloperAPI: This API may change across minor Ray releases.

on_start(worker_group: WorkerGroup, backend_config: BackendConfig)

Logic for starting this backend.

on_shutdown(worker_group: WorkerGroup, backend_config: BackendConfig)

Logic for shutting down the backend.

on_training_start(worker_group: WorkerGroup, backend_config: BackendConfig)

Logic ran right before training is started.

Session API is available at this point.