diff --git a/paddle/fluid/operators/sequence_ops/sequence_unpad_op.h b/paddle/fluid/operators/sequence_ops/sequence_unpad_op.h
index fe8ca41b698159a782547ce673a374d074d3b73d..70f26055b7cc0516b051fab2c8094752d5a9f9f1 100644
--- a/paddle/fluid/operators/sequence_ops/sequence_unpad_op.h
+++ b/paddle/fluid/operators/sequence_ops/sequence_unpad_op.h
@@ -33,7 +33,6 @@ class SequenceUnpadOpKernel : public framework::OpKernel<T> {
     auto* x_t = ctx.Input<LoDTensor>("X");
     auto* len_t = ctx.Input<LoDTensor>("Length");
     auto* out_t = ctx.Output<LoDTensor>("Out");
-    out_t->mutable_data<T>(ctx.GetPlace());
 
     const int64_t* seq_len_ptr = nullptr;
     if (platform::is_gpu_place(ctx.GetPlace())) {
@@ -67,6 +66,9 @@ class SequenceUnpadOpKernel : public framework::OpKernel<T> {
     }
     out_t->Resize(framework::make_ddim(out_dims_vec));
 
+    // after set the lod of output, allocate the memory
+    out_t->mutable_data<T>(ctx.GetPlace());
+
     int64_t padded_length = x_t->dims()[1];
     math::UnpaddingLoDTensorFunctor<DeviceContext, T>()(
         ctx.template device_context<DeviceContext>(), *x_t, out_t,