属性需要加下划线吗？is_dist_meta_

好的好的，疏忽了

欢哥，FillZeroForEmpty*会创建一个全0的Grad Tensor占有显存，但pp的非计算op前反向都不能占显存，需要控制一下进不进这个函数：

  bool rank_is_in_current_mesh = true;
  if (IsRunAutoParallel()) {
    auto mesh = std::static_pointer_cast<phi::distributed::DistTensor>(grads[{fwd_position}].impl())->dist_attr().process_mesh();
    rank_is_in_current_mesh = phi::distributed::IsCurRankInMesh(mesh);
  }
  if (rank_is_in_current_mesh) {{
    egr::EagerUtils::FillZeroForEmptyGradInput(&grads[{fwd_position}], input_metas[{fwd_position}]);
  }}

哦，这个是通过下面这种逻辑来控制的
if (tensor_meta.dims.size() != -1) {
auto tensor_with_zero =
paddle::experimental::full(phi::vectorize(tensor_meta.dims),
0.0,
tensor_meta.dtype,
grad_in_meta.GetPlace());
*(static_castphi::distributed::DistTensor*(in_grad->impl().get())
->unsafe_mutable_value()) =
*(static_castphi::DenseTensor*(tensor_with_zero.impl().get()));
}