o 1 i@s~ddlZddlmZmZddlmZddlmZedeedZ e e Z eddd e d e fd d Z edddd dZdS)N)OptionalTypeVar)get_train_fn_utils) PublicAPIT)boundalpha)Z stabilitydatareturncCs t|S)a]Broadcast small (<1kb) data from the rank 0 worker to all other workers. Serves as a barrier, meaning that all workers must call this method before the training function can continue. Example: .. testcode: from ray.train import get_context from ray.train.collective import broadcast_from_rank_zero from ray.train.torch import TorchTrainer def train_func(): ... if get_context().get_world_rank() == 0: data = {"some_key": "some_value"} else: data = None data = broadcast_from_rank_zero(data) ... trainer = TorchTrainer(train_func) trainer.fit() Args: data: The small (1kb) data to broadcast from the rank 0 worker to all other workers. Returns: The data broadcasted from the rank 0 worker. Raises: ValueError: If the data is too big. pickle.PicklingError: If the data is not pickleable. TypeError: If the data is not pickleable. )rbroadcast_from_rank_zero)r r l/home/app/PaddleOCR-VL-test/.venv_paddleocr/lib/python3.10/site-packages/ray/train/collective/collectives.pyr s 'r cCs tS)aCreate a barrier across all workers. All workers must call this method before the training function can continue. Example: .. testcode: from ray.train import get_context from ray.train.collective import barrier from ray.train.torch import TorchTrainer def train_func(): ... print(f"Rank {get_context().get_world_rank()} is waiting at the barrier.") barrier() print(f"Rank {get_context().get_world_rank()} has passed the barrier.") ... trainer = TorchTrainer(train_func) trainer.fit() )rbarrierr r r r r7s r)r N)loggingtypingrrZ/ray.train.v2._internal.execution.train_fn_utilsrZray.util.annotationsrobjectr getLogger__file__loggerr rr r r r s   )