diff --git a/paddle/fluid/operators/gru_op.h b/paddle/fluid/operators/gru_op.h
index 45c769ee37260bf912ebc848d58019557f4adc07..bcca992e2b426677e32d2c82e853d79534d114a6 100644
--- a/paddle/fluid/operators/gru_op.h
+++ b/paddle/fluid/operators/gru_op.h
@@ -144,7 +144,7 @@ class GRUGradKernel : public framework::OpKernel<T> {
         Tensor hidden_prev_grad_t = batch_hidden_grad.Slice(bstart_pre, bstart);
         gru_grad.prev_out_grad = hidden_prev_grad_t.data<T>();
       }
-
+      gru_value.output_value = nullptr;
       math::GRUUnitGradFunctor<DeviceContext, T>::compute(
           dev_ctx, gru_value, gru_grad, frame_size, cur_batch_size, active_node,
           active_gate, origin_mode);