TF_Status* status);

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>
void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,
    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>

  // custom device.
  //
  // Many devices will want to simply return an "unimplemented" status
  // here. This is the default behavior if `pack` is null when passed to
  // TFE_RegisterCustomDevice.
  TFE_TensorHandle* (*pack)(TFE_Context* context, TFE_TensorHandle** handles,
                            int num_handles, TF_Status* s,
                            void* device_info) = nullptr;