mindspore.communication.comm_func.P2POp
- class mindspore.communication.comm_func.P2POp(op, tensor, peer, group=None, tag=0, *, recv_dtype=None)[source]
Object for batch_isend_irecv input, to store information of
"isend"
and"irecv"
.Note
Allow pass-in recv shape rather than tensor when op is 'irecv'.
tensor will not be modified in-place by final result.
- Parameters
op (Union[str, function]) – Only string of
"isend"
and"irecv"
are allow. Or function ofcomm_func.isend
andcomm_func.irecv
are allow.tensor (Union[Tensor, Tuple(int)]) – tensor for sending/receiving or receive tensor shape when op is
"irecv"
.peer (int) – remote global rank for send/receive.
group (str, optional) – The communication group to work on. Default:
GlobalComm.WORLD_COMM_GROUP
, which means"hccl_world_group"
in Ascend, and"nccl_world_group"
in GPU.tag (int, optional) – currently not supported yet. default:
0
.
- Keyword Arguments
recv_dtype (mindspore.dtype, optional) – when tensor is a tuple shape, this arg will be used and has to be configured. default:
None
- Returns
P2POp Object.
- Raises
ValueError – when op is not string or function of 'isend' and 'irecv'.
TypeError – when tensor is not type of Tensor or Tuple.
NotImplementedError – when tag is not 0.
- Supported Platforms:
Ascend
Examples
>>> import numpy as np >>> import mindspore >>> from mindspore.communication.comm_func import P2POp, isend, irecv >>> from mindspore import Tensor >>> send_tensor = Tensor(1.) >>> send_op = P2POp('isend', send_tensor, 1) >>> send_op = P2POp(isend, send_tensor, 1) >>> recv_tensor = Tensor(0.) >>> recv_op = P2POp('irecv', recv_tensor, 0) >>> recv_op = P2POp(irecv, recv_tensor, 0) >>> recv_op = P2POp('irecv', (), 0, recv_dtype=mindspore.float32)