Created by: pkuyym

For variables in step_scopes->rbegin(), while_op would create a LoDTensor with same shape and same dtype as inside_tensor. However, the LoD is not inherited from insider_tensor, which results the output of while_grad_op being LoDTensor without LoD. However, in many situations, the LoD is necessary, for example:

x_tensor = fluid.layers.reorder_lod_tensor_by_rank(x, rank_table)
rnn = fluid.layers.DynamicRNN()
with rnn.block():
     x_step = rnn.step_input(x_tensor)

Input gradients of reorder_lod_tensor_by_rank_grad requires x being LoDTensor with LoD, however, the while_op_grad just dropped the LoD.

https://github.com/PaddlePaddle/Paddle/blob/c3e062e97565e5a790c826022c422c33a95d0b46/paddle/operators/while_op.cc#L180-L195

A possible fix is as following:

zero_op->Run(scope, dev_place);
auto* pg_tensor = scope.FindVar(pg_names[param_id])->GetMutable<framework::LoDTensor>();
pg_tensor->set_lod(inside_tensor.lod());

Inherit the LoD from inside_tensor explicitly.