TF_Function* TF_GraphToFunctionWithControlOutputs(
    const TF_Graph* fn_body, const char* fn_name,
    unsigned char append_hash_to_fn_name, int num_opers,
    const TF_Operation* const* opers, int ninputs, const TF_Output* inputs,
    int noutputs, const TF_Output* outputs, const char* const* output_names,
    int ncontrol_outputs, const TF_Operation* const* control_outputs,
    const char* const* control_output_names, const TF_FunctionOptions* opts,

  EXPECT_EQ(nodes_preserved.size(), num_values);
  EXPECT_EQ(list_total_size, storage_size);

  std::unique_ptr<char*[]> values(new char*[nodes_preserved.size()]);
  std::unique_ptr<size_t[]> lens(new size_t[nodes_preserved.size()]);
  std::unique_ptr<char[]> storage(new char[storage_size]);
  TF_GetNodesToPreserveList(c_item, values.get(), lens.get(),

// ParallelDevice.
void DeleteParallelDevice(void* device_info) {
  delete reinterpret_cast<NamedParallelDevice*>(device_info);
}

}  // namespace

void AllocateParallelDevice(const char* device_name,
                            const char* const* underlying_devices,
                            int num_underlying_devices,
                            TFE_CustomDevice* device, void** device_info) {

TFE_MonitoringNewCounterReader(const char* name);

// Reads the value of a counter that was created with 0 labels.
TF_CAPI_EXPORT extern int64_t TFE_MonitoringReadCounter0(
    TFE_MonitoringCounterReader*);

// Reads the value of specific cell of a counter that was created with 1 label.
TF_CAPI_EXPORT extern int64_t TFE_MonitoringReadCounter1(
    TFE_MonitoringCounterReader*, const char* label_value);

#ifdef __cplusplus

  // the operation, its `StartCancel` method will be called if op execution
  // fails on any device in order to cancel the others.
  void StartExecute(TFE_Context* context, const char* operation_name,
                    std::vector<TFE_TensorHandle*> inputs,
                    const TFE_OpAttrs* attributes, int expected_max_outputs,
                    CancellationManager& cancellation_manager,

  TFE_ContextSetServerDef(ctx, 0, serialized.data(), serialized.size(), status);
  EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  const char task0_name[] = "/job:localhost/replica:0/task:0/device:CPU:0";
  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.

  // if sanity checks on dtypes/metadata fail.
  absl::optional<std::vector<std::unique_ptr<ParallelTensor>>> Execute(
      TFE_Context* context, const std::vector<ParallelTensor*>& inputs,
      const char* operation_name, const TFE_OpAttrs* attributes,
      int expected_max_outputs, TF_Status* status) const;

  // A non-blocking version of `Execute`. After each call, `Join` must be called

namespace gradients {
namespace internal {
namespace {
using std::vector;
using tensorflow::TF_StatusPtr;
using tracing::TracingOperation;

class CppGradients
    : public ::testing::TestWithParam<std::tuple<const char*, bool, bool>> {
 protected:
  void SetUp() override {
    TF_StatusPtr status(TF_NewStatus());
    TF_SetTracingImplementation(std::get<0>(GetParam()), status.get());
    Status s = StatusFromTF_Status(status.get());

// and sets 'device_name' accordingly.
// `device_type` must be either "GPU" or "TPU".
bool GetDeviceName(TFE_Context* ctx, tensorflow::string* device_name,
                   const char* device_type);

// Create a ServerDef with the given `job_name` and add `num_tasks` tasks in it.
tensorflow::ServerDef GetServerDef(const tensorflow::string& job_name,
                                   int num_tasks);