all_reduce

paddle.distributed. all_reduce ( tensor, op=0, group=None, sync_op=True ) [source]

Reduce a tensor over all ranks so that all get the result. As shown below, one process is started with a GPU and the data of this process is represented by its group rank. The reduce operator is sum. Through all_reduce operator, each GPU will have the sum of the data from all GPUs.

all_reduce
Parameters
  • tensor (Tensor) – The input Tensor. It also works as the output Tensor. Its data type should be float16, float32, float64, int32, int64, int8, uint8 or bool.

  • op (ReduceOp.SUM|ReduceOp.MAX|ReduceOp.MIN|ReduceOp.PROD|ReduceOp.AVG, optional) – The operation used. Default value is ReduceOp.SUM.

  • group (Group, optional) – The group instance return by new_group or None for global default group.

  • sync_op (bool, optional) – Wether this op is a sync op. Default value is True.

Returns

Return a task object.

Examples

>>> 
>>> import paddle
>>> import paddle.distributed as dist

>>> dist.init_parallel_env()
>>> if dist.get_rank() == 0:
...     data = paddle.to_tensor([[4, 5, 6], [4, 5, 6]])
>>> else:
...     data = paddle.to_tensor([[1, 2, 3], [1, 2, 3]])
>>> dist.all_reduce(data)
>>> print(data)
>>> # [[5, 7, 9], [5, 7, 9]] (2 GPUs)