ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Create a tensor on the first parallel device
  TensorHandlePtr value_one(FloatTensorHandle(1., status.get()));
  TensorHandlePtr value_two(FloatTensorHandle(2., status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  std::array<TFE_TensorHandle*, 2> components{value_one.get(), value_two.get()};

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

  TensorHandlePtr value_one(FloatTensorHandle(3., status.get()));
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TensorHandlePtr value_two(FloatTensorHandle(-2., status.get()));
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    ASSERT_EQ(underlying_devices[1], second_device);
  }

  // Add a parallel tensor with different values on each device to the variable.
  {
    TensorHandlePtr value_one(FloatTensorHandle(3., status.get()));
    TensorHandlePtr value_two(FloatTensorHandle(-2., status.get()));
    std::array<TFE_TensorHandle*, 2> components{value_one.get(),
                                                value_two.get()};
    TensorHandlePtr combined_value =

  // The dtype of the variable's buffer (input dtype for assignments, output
  // dtype of read operations).
  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);

  const std::vector<int64_t>* shape;
  TF_ASSERT_OK(unknown_length_vector->Shape(&shape));
  EXPECT_THAT(*shape, ElementsAre(-1));

  TensorHandlePtr scalar = FloatTensorHandle(2., status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  two_vector = VectorFloatTensorHandle({3., 4.}, status.get());