Differences with torch.distributed.scatter

torch.distributed.scatter

torch.distributed.scatter(
    tensor,
    scatter_list=None,
    src=0,
    group=None,
    async_op=False
)

For more information, see torch.distributed.scatter。

mindspore.communication.comm_func.scatter_tensor

mindspore.communication.comm_func.scatter_tensor(tensor, src=0, group=GlobalComm.WORLD_COMM_GROUP)

For more information, see mindspore.communication.comm_func.scatter_tensor。

Differences

API function of MindSpore is not consistent with that of PyTorch.

PyTorch：tensor will store the scattered result, the tensors to be gatherd scatter_list, the rank(global rank) src of the process, the communication group to work on, and the async op flag async_op. The process will scatter the tensors and store in tensor. The return is a async work handle if async_op is True, otherwise is None. The async op flag async_op and the working device lists device_ids are not supported.

MindSpore: The inputs contains the tensor to be scattered, the rank(global rank) src of the process, the communication group to work on, and it will return the tensor. The dimension 0 of data is equal to the dimension of input tensor divided by src, and the other dimension keep the same.