0维Tensor，数据并行梯度处理报错

[wangguan1995]

bug描述 Describe the Bug

最小复现

# https://github.com/PaddlePaddle/PaddleScience/blob/develop/ppsci/arch/activation.py
import paddle
import paddle.nn as nn
import paddle.distributed as dist
from paddle.distributed import fleet
import paddle.nn.functional as F
from paddle.distributed.fleet.utils import hybrid_parallel_util as hpu
dist.init_parallel_env()

class Swish(nn.Layer):
    def __init__(self, beta: float = 1.0):
        super().__init__()
        self.linear = nn.Linear(5,1)
        self.beta = self.create_parameter(
            shape=[],
            default_initializer=nn.initializer.Constant(beta),
        )

    def forward(self, x):
        x = self.linear(x)
        return x * F.sigmoid(self.beta * x)

if __name__ == '__main__':
    model = Swish()

    if dist.get_world_size() > 1:
        fleet.init(is_collective=True)
        model = fleet.distributed_model(model)


        y = model(paddle.to_tensor([-2., -1., 0., 1., 2.]))
        loss = y - paddle.to_tensor([-1., 0., 1., 2., 3.])
        loss.backward()

        # fuse + allreduce manually before optimization if use DDP + no_sync
        # details in https://github.com/PaddlePaddle/Paddle/issues/48898#issuecomment-1343838622
        p = list(model.parameters())
        for i, (name, param) in enumerate(model.named_parameters()):
            print(f"Layer: {name} | Size: {param.shape}, {p[i].shape}")
            print(param.trainable)
            print(param._grad_ivar() is not None) # true after backward

        hpu.fused_allreduce_gradients(list(model.parameters()), None)
        print(model(paddle.to_tensor([-2., -1., 0., 1., 2.])))
    else:
        raise NotImplementedError

其他补充信息 Additional Supplementary Information

paddle commit id :
e2894ad

[LokeZhou]

你好，本地暂时没有复现错误，后面持续看看