}

void TFE_HostAddressSpace(TFE_Context* ctx, TF_Buffer* buf) {
  auto address_space = tensorflow::DeviceNameUtils::AddressSpace(
      tensorflow::unwrap(ctx)->HostCPUParsedName());
  auto str = tensorflow::DeviceNameUtils::ParsedNameToString(address_space);
  void* data = tensorflow::port::Malloc(str.length());
  str.copy(static_cast<char*>(data), str.length(), 0);
  buf->data = data;

  // Update the vector with information from `input_tensors` if provided.
  if (input_tensors != nullptr) {
    // Note that we take the address of the elements in `all_input_tensors`
    // below. Allocate enough space so that no reallocation happens, which will
    // make the pointers invalid.
    all_input_tensors.reserve(num_inputs);
    for (int i = 0; i < num_inputs; ++i) {
      if (input_tensors[i] == nullptr) continue;

    *   Added int16x8 support for the built-in op `exp`
    *   Added int16x8 support for the built-in op `mirror_pad`
    *   Added int16x8 support for the built-in ops `space_to_batch_nd` and `batch_to_space_nd`
    *   Added 16-bit int type support for built-in op `less`, `greater_than`, `equal`
    *   Added 8-bit and 16-bit support for `floor_div` and `floor_mod`.

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) {
    call_state_.emplace(nullptr, false);
  }

  virtual ~ForwardAccumulator() {