reduce_scatter

paddle.distributed.communication.stream. reduce_scatter ( tensor, tensor_or_tensor_list, op=0, group=None, sync_op=True, use_calc_stream=False ) [source]

Reduce, then scatter a tensor (or a tensor list) across devices.

Parameters

  • tensor (Tensor) – The output tensor on each rank. The result will overwrite this tenor after communication. Support float16, float32, float64, int32, int64, int8, uint8 or bool as the input data type.

  • tensor_or_tensor_list (Union[Tensor, List[Tensor]]) – The input to scatter. If it is a tensor, it should be correctly-sized. If it is a list, it should contain correctly-sized tensors.

  • op (ReduceOp.SUM|ReduceOp.MAX|ReduceOp.MIN|ReduceOp.PROD, optional) – The reduction used. If none is given, use ReduceOp.SUM as default.

  • group (Group, optional) – Communicate in which group. If none is given, use the global group as default.

  • sync_op (bool, optional) – Indicate whether the communication is sync or not. If none is given, use true as default.

  • use_calc_stream (bool, optional) – Indicate whether the communication is done on calculation stream. If none is given, use false as default. This option is designed for high performance demand, be careful to turn it on except you are clearly know its meaning.

Returns

Return a task object.

Warning

This API only supports the dygraph mode now.

Examples

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

>>> dist.init_parallel_env()
>>> if dist.get_rank() == 0:
...     data1 = paddle.to_tensor([0, 1])
...     data2 = paddle.to_tensor([2, 3])
>>> else:
...     data1 = paddle.to_tensor([4, 5])
...     data2 = paddle.to_tensor([6, 7])
>>> dist.stream.reduce_scatter(data1, [data1, data2])
>>> out = data1.numpy()
>>> print(out)
>>> # [4, 6]  (2 GPUs, out for rank 0)
>>> # [8, 10] (2 GPUs, out for rank 1)