// This method is used to add a new input edge to 'dst', which must be a While
// op. The While op's "T" attribute must have already been updated to include
// the new edge. This is used to construct tf.while_loop gradients.
TF_CAPI_EXPORT extern void TF_AddWhileInputHack(TF_Graph* graph,
                                                TF_Output new_src,
                                                TF_Operation* dst,

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override;
};
// Shorthand for registry->Register(op, new NotDifferentiableGradientFunction)
Status RegisterNotDifferentiable(GradientRegistry* registry, const string& op);
}  // namespace gradients
}  // namespace tensorflow

// read current values from streamz counters defined in other modules.
//
// The code under test will have created streamz counters like this:
// auto* streamz = tensorflow::monitoring::Counter<1>::New("name",
// "description", "label");
// and then incremented that counter for various values of label:
// streamz->GetCell("label-value")->IncrementBy(1);
//
// The test code can then read and test the value of that counter:

  // `status` checked.
  void Destroy(TFE_Context* context, TF_Status* status);

  // Reads from the variable.
  TensorHandlePtr Read(TFE_Context* context, TF_Status* status);
  // Assigns a new value to the variable.
  void Assign(TFE_Context* context, TFE_TensorHandle* value, TF_Status* status);
  // Adds `value` to the existing value of the variable.
  void AssignAdd(TFE_Context* context, TFE_TensorHandle* value,

TF_CAPI_EXPORT extern uint64_t TF_NowSeconds(void);

// Populates a TF_ThreadOptions struct with system-default values.
TF_CAPI_EXPORT extern void TF_DefaultThreadOptions(TF_ThreadOptions* options);

// Returns a new thread that is running work_func and is identified
// (for debugging/performance-analysis) by thread_name.
//
// The given param (which may be null) is passed to work_func when the thread

  template <typename... LabelDesc>
  TFE_MonitoringCounter(const char* name, const char* description,
                        LabelDesc&&... label) {
    counter = absl::WrapUnique(tensorflow::monitoring::Counter<NumLabels>::New(
        name, description, label...));
  }

  std::unique_ptr<tensorflow::monitoring::Counter<NumLabels>> counter;
};

struct TFE_MonitoringCounter0 : TFE_MonitoringCounter<0> {

                                                         TF_Status* status);

// Create a new TFE_TensorHandle with the same contents as 'h' but placed
// in the memory of the device name 'device_name'.
// If source and destination are the same device, then this creates a new handle
// that shares the underlying buffer. Otherwise, it currently requires at least

class ImmediateExecutionDistributedManager {
 public:
  virtual ~ImmediateExecutionDistributedManager() {}

  // Set up distributed execution environment on local and remote tasks.
  // When `reset_context` is true, initialize new cluster context state based
  // on cluster configurations provided in `server_def`; otherwise, update
  // existing context state with the provided `server_def`. Contexts created

  template <typename G>
  Cleanup(Cleanup<G>&& src)  // NOLINT
      : released_(src.is_released()), f_(src.release()) {}

  // Assignment to a Cleanup object behaves like destroying it
  // and making a new one in its place, analogous to unique_ptr
  // semantics.
  Cleanup& operator=(Cleanup&& src) {  // NOLINT
    if (!released_) f_();
    released_ = src.released_;
    f_ = src.release();
    return *this;
  }

  }

  // We may need to allocate zero inputs for trainable dtypes we don't have JVPs
  // for. Make sure they get cleaned up.
  std::vector<Gradient*> new_zeros;
  auto delete_new_zeros = gtl::MakeCleanup([&new_zeros, this] {
    for (Gradient* tensor : new_zeros) {
      this->vspace_.DeleteGradient(tensor);
    }
  });
  std::vector<Gradient*> in_grads;
  in_grads.reserve(input_tensors.size());