std::vector<PartialTensorShape> shapes;
    shapes.reserve(num_values);
    for (int i = 0; i < num_values; ++i) {
      if (num_dims[i] < 0) {
        shapes.emplace_back();
      } else {
        shapes.emplace_back(ArraySlice<int64_t>(
            reinterpret_cast<const int64_t*>(dims[i]), num_dims[i]));
      }
    }
    op_->node_builder.Attr(attr_name, shapes);
    return absl::OkStatus();
  }

      // One of the inputs we're trying to pack is on a custom device. We'll let
      // the first custom device we see handle all of the packing.
      auto* custom_device_handle =
          tensorflow::down_cast<tensorflow::CustomDeviceTensorHandle*>(
              unwrapped_handle);
      tensorflow::ImmediateExecutionTensorHandle* result;
      status->status = custom_device_handle->device()->Pack(

  }
}

/// Appends a trailing slash if the name doesn't already have one.
static void MaybeAppendSlash(std::string* name) {
  if (name->empty())
    *name = "/";
  else if (name->back() != '/')
    name->push_back('/');
}

// A helper function to actually read the data from GCS.
static int64_t LoadBufferFromGCS(const std::string& path, size_t offset,

  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  const int num_devices = TF_DeviceListCount(devices);
  for (int i = 0; i < num_devices; ++i) {
    device_names.emplace_back(TF_DeviceListName(devices, i, status));
    EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  }
  TF_DeleteDeviceList(devices);
  TF_DeleteStatus(status);
  return device_names;

        // GradientTape which uses ones.
        Gradient* zero = input_tensors[target_index].ZerosLike();
        new_zeros.push_back(zero);
        in_grads.push_back(zero);
      } else {
        in_grads.push_back(nullptr);
      }
    } else {
      in_grads.push_back(current_grad->second);
    }
  }

  // Avoid infinite recursion. Whichever forward function we run, it'll end up

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  std::vector<TensorHandlePtr> vector_handles;
  vector_handles.reserve(2);
  vector_handles.push_back(std::move(two_vector));
  vector_handles.push_back(std::move(three_vector));
  std::unique_ptr<ParallelTensor> unknown_length_vector =
      ParallelTensor::FromTensorHandles(
          parallel_device, std::move(vector_handles), status.get());

    all_input_tensors.reserve(num_inputs);
    for (int i = 0; i < num_inputs; ++i) {
      if (input_tensors[i] == nullptr) continue;
      all_input_tensors.emplace_back();
      Tensor& input_tensor = all_input_tensors.back();
      status->status = TF_TensorToTensor(input_tensors[i], &input_tensor);
      if (!status->status.ok()) return;
      input_tensors_vector[i] = &input_tensor;
    }
  }

        return result;
      }
      components.emplace_back(TFE_TensorHandleCopySharingTensor(
          absl::get<TFE_TensorHandle*>(inputs[i]), status));
    }
    std::vector<MaybeParallelTensorOwned> result_content;
    result_content.reserve(1);
    result_content.push_back(ParallelTensor::FromTensorHandles(
        parallel_device, std::move(components), status));

    components.reserve(underlying_devices_.size());
    for (int j = 0; j < underlying_devices_.size(); ++j) {
      components.push_back(std::move(per_device_output_tensors[j][i]));
    }
    if (expected_output_shapes[i].IsFullyDefined()) {
      per_device_outputs.push_back(ParallelTensor::FromTensorHandles(
          *this, std::move(components),
          absl::Span<const int64_t>(expected_output_shapes[i].dim_sizes()),

    for (int i = 0; i < noutputs; ++i) {
      output_names_vec.push_back(string(output_names[i]));
    }
  }

  // Process control output names.
  std::vector<string> control_output_names_vec;
  if (control_output_names) {
    control_output_names_vec.reserve(ncontrol_outputs);
    for (int i = 0; i < ncontrol_outputs; ++i) {
      control_output_names_vec.push_back(string(control_output_names[i]));
    }
  }