// must point to an array of length at least `num_values`.
//
// The elements of values will point to addresses in `storage` which must be at
// least `storage_size` bytes in length.  `num_values` and `storage` can be
// obtained from TF_GetNodesToPreserveSize
//
// Fails if storage_size is too small to hold the requested number of strings.
TF_CAPI_EXPORT extern void TF_GetNodesToPreserveList(

TF_CAPI_EXPORT extern int64_t TFE_TensorDebugInfoOnDeviceDim(
    TFE_TensorDebugInfo* debug_info, int dim_index);

// Description of the TensorFlow op to execute.
//
// Assumes that the provided 'ctx' outlives the returned TFE_Op, i.e.,
// TFE_DeleteOp() is called before TFE_DeleteContext().
//
// Very similar to TF_OperationDescription with some differences:
// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,

struct TF_AttrBuilder : public tensorflow::AttrBuilder {
  using tensorflow::AttrBuilder::AttrBuilder;
  // The string buffers to make sure that any `attr_name` we pass into
  // `builder->Set()` will outlive the subsequent
  // `TF_AttrBuilderCheckCanRunOnDevice()` call(s) on the same `builder`.
  std::set<std::string> attr_names;
};

TF_AttrBuilder* TF_NewAttrBuilder(const char* op_name) {

//
// Not thread-safe.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class ForwardAccumulator {
 public:
  // Does not take ownership of `vspace`, which must outlive the
  // ForwardAccumulator.
  explicit ForwardAccumulator(
      const VSpace<Gradient, BackwardFunction, TapeTensor>& vspace,
      bool use_batch)
      : vspace_(vspace), use_batch_(use_batch) {

  // re-use a thread.
  std::vector<std::unique_ptr<DeviceThread>> device_threads_;
  // A cancellation manager to use if the caller does not provide one. When ops
  // are executed asynchronously this must outlive the queued op, so it can't be
  // function-local to Execute.
  mutable std::unique_ptr<CancellationManager> default_cancellation_manager_;
};