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;
}

    // 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());

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 {

  // The resulting TensorHandle owns an opaque pointer to "device memory", which
  // for a ParallelDevice is really a ParallelTensor. When the TensorHandle is
  // deleted, it will call ParallelTensorDeallocator to free the struct.
  ParallelTensor* t_released = t.release();
  TFE_CustomDeviceTensorHandleMethods handle_methods;
  handle_methods.num_dims = &ParallelTensorNumDims;
  handle_methods.dim = &ParallelTensorDim;

//     probe for gadgets inside core TensorFlow. We can even protect the area
//     of memory where the copies reside to not allow any more writes to it
//     after all copies are created.
template <typename T>
static std::unique_ptr<const T> CopyToCore(const T* plugin_ops,
                                           size_t plugin_size) {
  if (plugin_ops == nullptr) return nullptr;

  Define(-1, {}, {feed}, {neg}, {"negated_num"});

  // Use, run, and verify
  TF_Operation* func_feed = Placeholder(host_graph_, s_);
  TF_Operation* func_op = Use({func_feed});
  Run({{func_feed, Int32Tensor(3)}}, func_op, -3);
  VerifyFDef({"neg"}, {{"feed", DT_INT32}}, {{"negated_num", DT_INT32}},
             {{"feed", "neg:0"}, {"neg:y:0", "negated_num"}}, {});
}

    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);