handle_methods.summarize = &ParallelTensorSummarize;
  return TensorHandlePtr(TFE_NewCustomDeviceTensorHandle(
      context, parallel_device_name.c_str(), t_released->dtype(), t_released,
      handle_methods, status));
}

// For TFE_CustomDevice::copy_tensor_to_device in the parallel device
// registration.
//
// Since this function is used to satisfy the TFE_CustomDevice C API,

     */
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Get transpose attrs
    bool t_a;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_a", &t_a));

    bool t_b;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_b", &t_b));

    // Conj each input
    AbstractTensorHandle* conj_output;
    std::string name = "Conj_A_MatMul_Grad";
    TF_RETURN_IF_ERROR(

  TF_DeleteTensor(a);
  TF_DeleteTensor(b);
}

TEST(CAPI, TestFromProto) {
  Tensor t_cc(DT_FLOAT, TensorShape({2, 3}));
  t_cc.flat<float>().setConstant(1.0);
  tensorflow::TensorProto t_proto;
  t_cc.AsProtoField(&t_proto);

  TF_Buffer* t_buffer = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(t_proto, t_buffer));

  const int num_bytes = 6 * sizeof(float);
  float* values =