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;

  //   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_);

//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either
// need to return something here or address these use-cases one by one.
TFE_TensorHandle* CopyTensorFromParallelDevice(TFE_Context* context,
                                               TFE_TensorHandle* tensor,
                                               const char* target_device_name,

  const char task1_name[] = "/job:localhost/replica:0/task:1/device:CPU:0";
  const char task2_name[] = "/job:localhost/replica:0/task:2/device:CPU:0";

  // Create one variable per task.
  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/platform/errors.h"

namespace tensorflow {

// Checks that all schemes provided by a plugin are valid.
// TODO(b/139060984): More validation could be done here, based on supported
// charset, maximum length, etc. Punting it for later.
static Status ValidateScheme(const char* scheme) {
  if (scheme == nullptr)
    return errors::InvalidArgument(

// TODO(vnvo2409): We could do some cleanups like `return TF_SetStatus`.
// TODO(vnvo2409): Refactor the filesystem implementation when
// https://github.com/googleapis/google-cloud-cpp/issues/4482 is done.
GCSFile::GCSFile(google::cloud::storage::Client&& gcs_client)
    : gcs_client(gcs_client), block_cache_lock() {
  const char* append_mode = std::getenv(kAppendMode);

                             params->body_graph, s_, "add1");
    ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
    TF_Operation* one = ScalarConst(1, params->body_graph, s_);
    ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
    TF_Operation* add2 = Add(add1, one, params->body_graph, s_, "add2");
    ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
    params->body_outputs[0] = {add2, 0};