TF_Output output{dequeue_op, 0};
  TF_Tensor* ret;
  TF_SessionRun(session, /*run_options*/ nullptr,
                // input related parameters
                /*inputs*/ nullptr, /*input_values*/ nullptr, /*ninputs*/ 0,
                // output related parameters
                /*outputs*/ &output, /*output_values*/ &ret,
                /*noutputs*/ 1,
                /*targets*/ nullptr, /*ntargets*/ 0,

  while (*pos < static_cast<size_t>(graph->graph.num_node_ids())) {
    Node* node = graph->graph.FindNodeId(*pos);
    // FindNodeId() returns nullptr for nodes that have been deleted.
    // We aren't currently allowing nodes to be deleted, but it is safer
    // to still check.
    if (node != nullptr) return ToOperation(node);
    *pos += 1;
  }

  // No more nodes.
  return nullptr;
}

namespace {

// Managing context for the DLManagedTensor, will manage the lifetime of
// DLManagedTensor. When calling DLManagedTensor::deleter, it will notify the
// original framework of destruction, and this context will be deleted also.
struct TfDlManagedTensorCtx {
  TensorReference reference;
  std::vector<int64_t> shape;
  std::vector<int64_t> strides;
  DLManagedTensor tensor;

  void MarkAsResult(AbstractTensorHandle* gradient) const override;

  void DeleteGradient(AbstractTensorHandle* gradient) const override;

 private:
  // The context where the aggregation op `Add` is to be created.
  AbstractContext* ctx_;
};

// Returns the number of elements in the gradient tensor.
int64_t TapeVSpace::NumElements(AbstractTensorHandle* tensor) const {

  TF_Output add_inputs[2] = {{l, 0}, {r, 0}};
  TF_AddInputList(desc, add_inputs, 2);
  TF_AddControlInput(desc, ctrl_op);
  return TF_FinishOperation(desc, s);
}

// If `op_device` is non-empty, set the created op on that device.
void BinaryOpHelper(const char* op_name, TF_Operation* l, TF_Operation* r,
                    TF_Graph* graph, TF_Status* s, const char* name,

    // would otherwise deadlock).
    if (TF_GetCode(status) != TF_OK &&
        (first_bad_status == nullptr
         // Prefer propagating non-cancellation related statuses to avoid
         // shadowing the original failure.
         || TF_GetCode(first_bad_status.get()) == TF_CANCELLED)) {
      first_bad_status.reset(TF_NewStatus());

  absl::MutexLock l(&block->mu);
  TF_SetStatus(status, TF_OK, "");
  while (true) {
    switch (block->state) {
      case FetchState::ERROR:
        // TF_FALLTHROUGH_INTENDED
      case FetchState::CREATED:
        block->state = FetchState::FETCHING;
        block->mu.Unlock();  // Release the lock while making the API call.
        block->data.clear();
        block->data.resize(block_size_, 0);

    if (forward_input) {
      // Since the input was forwarded, we released the input handle right away
      // and hence expect the input to be forwarded to the return tensor.
      EXPECT_EQ(orig_ptr, TF_TensorData(t));
    } else {
      // In async mode we expect forwarding to work without releasing the input
      // handle since by the time the kernel is executed we have released the
      // handle in the client code.

  //
  // 1. Create a variable on `remote_device`, using `ctx_0`.
  TFE_TensorHandle* handle_0 =
      CreateVariable(ctx_0, 1.2, remote_device, /*variable_name=*/"var");

  // 2. Wait for `var` to be created and initialized on the worker.
  TF_Status* status = TF_NewStatus();
  TFE_ContextAsyncWait(ctx_0, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteStatus(status);

      TF_FinalizeFunction(graph_ctx, add_outputs, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  // Note: TF_OutputList does not own the underlying AbstractTensors, those
  // need to be deleted explicitly.
  TF_DeleteAbstractTensor(TF_OutputListGet(add_outputs, 0));

  // Build eager context.
  TFE_ContextOptions* opts = TFE_NewContextOptions();
  TFE_ContextOptionsSetTfrt(opts, std::get<1>(GetParam()));