ParallelDevice parallel_device(std::move(devices));
  TensorHandlePtr two_vector = VectorFloatTensorHandle({3., 4.}, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TensorHandlePtr three_vector =
      VectorFloatTensorHandle({5., 6., 7.}, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  TF_DataType type_;
};

// Creates a TFE_TensorHandle with value `v`.
TensorHandlePtr FloatTensorHandle(float v, TF_Status* status);

// Creates a rank-one TFE_TensorHandle with value `v`.
TensorHandlePtr VectorFloatTensorHandle(const std::vector<float>& v,
                                        TF_Status* status);

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>

  std::vector<float> size_two_value{1., 2.};
  std::vector<float> size_three_value{1., 2., 3.};
  TensorHandlePtr size_two(
      VectorFloatTensorHandle(size_two_value, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  TensorHandlePtr size_three(
      VectorFloatTensorHandle(size_three_value, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

                   nullptr),
      TF_DeleteTensor);
  return TensorHandlePtr(TFE_NewTensorHandle(tensor.get(), status));
}

// Creates a rank-one TFE_TensorHandle with value `v`.
TensorHandlePtr VectorFloatTensorHandle(const std::vector<float>& v,
                                        TF_Status* status) {
  const int num_bytes = v.size() * sizeof(float);
  float* values = new float[v.size()];
  memcpy(values, v.data(), num_bytes);