// reference counts drop to zero.
void ParallelTensorDeallocator(void* data) {
  delete reinterpret_cast<ParallelTensor*>(data);
}

// Used as an argument to TFE_NewCustomDeviceTensorHandle, for computing the
// number of dimensions of a parallel tensor.
int ParallelTensorNumDims(void* data, TF_Status* status) {
  const std::vector<int64_t>* shape;
  Status s = reinterpret_cast<ParallelTensor*>(data)->Shape(&shape);
  if (!s.ok()) {

        context.get(), components, device_name, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_INVALID_ARGUMENT)
        << TF_Message(status.get());
  }

  {
    // Try to extract the wrong number of components from a parallel tensor
    std::array<TFE_TensorHandle*, 1> correct_components{value_one.get()};
    TensorHandlePtr combined_value = CreatePerDeviceValues(