TF_Status* status) const;

  void AsyncWait(TFE_Context* context, TF_Status* status) const;

  // Device strings for component devices that only include a
  // worker/task/replica if any of those differ across components. Useful for
  // printing debug messages.
  std::vector<std::string> SummarizeDeviceNames() const;

 private:
  // A sequence of device names, indicating which devices replicated operations

                                              TF_Tensor* to, TF_Status* status);

// --------------------------------------------------------------------------
// Used to return strings across the C API. The caller does not take ownership
// of the underlying data pointer and is not responsible for freeing it.
typedef struct TF_StringView {
  const char* data;
  size_t len;
} TF_StringView;

  }

  // 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());

  // Returns number of elements across all dimensions.
  virtual Status NumElements(int64_t* num_elements) const = 0;
  // Returns size of specified dimension
  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.

// the datatype for boolean tensors.
#ifndef TF_Bool
#define TF_Bool unsigned char
#endif  // TF_Bool

// Macro used to calculate struct size for maintaining ABI stability across
// different struct implementations.
#ifndef TF_OFFSET_OF_END
#define TF_OFFSET_OF_END(TYPE, MEMBER) \
  (offsetof(TYPE, MEMBER) + sizeof(((TYPE *)0)->MEMBER))
#endif  // TF_OFFSET_OF_END

// allow us to trace the data dependencies between operations and hence compute
// gradients.
//
// `ZerosLike` is not expected to be called and returns a nullptr. The creation
// of default zeros grads is handled by the `DefaultGradientFunction` registered
// for each op.
// TODO(srbs): We need to define `ZerosLike` here to keep the compiler happy.
// Figure out a way to avoid this.