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());
  std::array<TFE_TensorHandle*, 2> in_components{value_one.get(),

      "important to you");
  return EmptyWhileParams();
#else
  if (ninputs == 0) {
    status->status =
        InvalidArgument("TF_NewWhile() must be passed at least one input");
    return EmptyWhileParams();
  }

  TF_Graph* cond_graph = TF_NewGraph();
  TF_Graph* body_graph = TF_NewGraph();
  cond_graph->parent = g;
  cond_graph->parent_inputs = inputs;

  auto collective_executor_handle = context->GetCollectiveExecutorHandle();
  tensorflow::Notification done;
  collective_executor_handle->get()->remote_access()->CheckPeerHealth(
      task, timeout_in_ms, [&done, status](const Status& s) {
        status->status = s;
        done.Notify();
      });
  done.WaitForNotification();
}

TF_ShapeAndTypeList* TF_NewShapeAndTypeList(int num_items) {

}

TEST_P(CppGradients, TestMatMulGrad) {
  // TODO(vnvo2409): Figure out why `gradient_checker` does not work very
  // well with `MatMul` and remove `TestMatMul*` in
  // `mnist_gradients_test` when done.
  GTEST_SKIP();

  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f};
  int64_t A_dims[] = {3, 3};
  AbstractTensorHandlePtr A;
  {
    AbstractTensorHandle* A_raw;

    const char* childpath;
    ASSERT_TRUE(TF_StringStreamNext(children, &childpath));
    ASSERT_EQ(::tensorflow::string(childpath), "somefile.txt");
    // There should only be one file in this directory.
    ASSERT_FALSE(TF_StringStreamNext(children, &childpath));
    ASSERT_EQ(childpath, nullptr);
    TF_StringStreamDone(children);

    TF_FileStatistics stats;

  std::unique_ptr<TFE_TensorHandle, decltype(&TFE_DeleteTensorHandle)> one(
      TestScalarTensorHandle(context.get(), 111.f), TFE_DeleteTensorHandle);
  op.reset(TFE_NewOp(context.get(), "AssignVariableOp", status.get()));
  TFE_OpSetAttrType(op.get(), "dtype", TF_FLOAT);
  TFE_OpAddInput(op.get(), var_handle, status.get());
  TFE_OpAddInput(op.get(), one.get(), status.get());
  TFE_OpSetDevice(op.get(), name, status.get());

  values[0] = v;
  return TF_NewTensor(TF_FLOAT, nullptr, 0, values, num_bytes,
                      &FloatDeallocator, nullptr);
}

// All the *Helper methods are used as a workaround for the restrictions that
// one cannot call ASSERT_* methods in non-void-returning functions (when
// exceptions are disabled during compilation)
void PlaceholderHelper(TF_Graph* graph, TF_Status* s, const char* name,

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

  TensorHandlePtr negative_one_cpu(FloatTensorHandle(3., status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  components[0] = negative_one_cpu.get();
  components[1] = negative_one_cpu.get();
  TensorHandlePtr first_negative_one = CreatePerDeviceValues(
      context.get(), components, first_device_name, status.get());

  }

  // 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 =

  //   Outputs:
  std::vector<TensorHandlePtr> op_outputs_ TF_GUARDED_BY(execution_mutex_);
  // TF_Status is an incomplete type and so can't be stack allocated. To avoid
  // unnecessary allocations each Execute call, we keep one heap-allocated
  // version for the thread.
  StatusPtr status_ TF_GUARDED_BY(execution_mutex_);

  const std::string device_;
  ExecutorPtr executor_ TF_GUARDED_BY(execution_mutex_);