// Method to copy a tensor from the custom device to a target device.
  TFE_TensorHandle* (*copy_tensor_from_device)(TFE_Context* context,
                                               TFE_TensorHandle* tensor,
                                               const char* target_device_name,
                                               TF_Status* status,

  in_grads.reserve(input_tensors.size());
  for (int target_index = 0; target_index < input_tensors.size();
       ++target_index) {
    const auto current_grad =
        accumulated_gradients_.find(input_tensors[target_index].GetID());
    if (current_grad == accumulated_gradients_.end()) {
      if (IsDtypeTrainable(input_tensors[target_index].GetDType())) {