TensorHandlePtr first_negative_one = CreatePerDeviceValues(
      context.get(), components, first_device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  components[0] = first_negative_one.get();
  components[1] = negative_one_cpu.get();
  TensorHandlePtr second_negative_one = CreatePerDeviceValues(
      context.get(), components, second_device_name, status.get());

    std::array<TensorHandlePtr, num_replicas>* components, TF_Status* status);

// Helper to pack `num_replicas` TFE_TensorHandles into one parallel handle.
template <std::size_t num_replicas>
TensorHandlePtr CreatePerDeviceValues(
    TFE_Context* context,
    const std::array<TFE_TensorHandle*, num_replicas>& components,
    const char* device, TF_Status* status);

TensorHandlePtr Multiply(TFE_Context* context, TFE_TensorHandle* first,

    std::array<TFE_TensorHandle*, 2> components{initial_value_cpu.get(),
                                                initial_value_cpu.get()};
    TensorHandlePtr initial_value =
        CreatePerDeviceValues(context, components, device_name, status.get());
    variable->Assign(context, initial_value.get(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  }

  std::array<TFE_TensorHandle*, 2> in_components{value_one.get(),
                                                 value_two.get()};
  TensorHandlePtr combined_value = CreatePerDeviceValues(
      context.get(), in_components, device_name, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Loop to make synchronization failures more deterministic