@ -66,7 +66,7 @@ void FetchOpHandle::RunImpl() {
auto &t = var->Get<framework::LoDTensor>();
if (platform::is_gpu_place(t.place())) {
#ifdef PADDLE_WITH_CUDA
TensorCopy(t, cpu, *dev_ctxes_[t.place()], &tensors_[i]);
TensorCopy(t, cpu, *dev_ctxes_[t.place()], &tensors_[i], true);
dev_ctxes_.at(t.place())->Wait();
#endif
} else {