return result;
  }
  std::vector<ParallelTensor*> parallel_inputs;
  std::vector<std::unique_ptr<ParallelTensor>> implicitly_broadcast_tensors;
  parallel_inputs.reserve(inputs.size());
  implicitly_broadcast_tensors.reserve(inputs.size());  // not tight
  for (const auto& input : inputs) {
    if (absl::holds_alternative<TFE_TensorHandle*>(input)) {
      if (operation_name == std::string("_EagerConst")) {

  std::vector<std::unique_ptr<ParallelTensor>> per_device_outputs;
  per_device_outputs.reserve(first_op_output_count);
  for (int i = 0; i < first_op_output_count; ++i) {
    std::vector<TensorHandlePtr> components;
    components.reserve(underlying_devices_.size());
    for (int j = 0; j < underlying_devices_.size(); ++j) {

  if (output_names) {
    output_names_vec.reserve(noutputs);
    for (int i = 0; i < noutputs; ++i) {
      output_names_vec.push_back(string(output_names[i]));
    }
  }

  // Process control output names.
  std::vector<string> control_output_names_vec;
  if (control_output_names) {
    control_output_names_vec.reserve(ncontrol_outputs);
    for (int i = 0; i < ncontrol_outputs; ++i) {

}

inline Status AbstractOperation::SetAttrStringList(
    const char* attr_name, absl::Span<string const> values) {
  std::vector<const char*> raw_strs;
  std::vector<size_t> lengths;
  raw_strs.reserve(values.size());
  lengths.reserve(values.size());
  for (const auto& s : values) {
    raw_strs.emplace_back(s.data());
    lengths.emplace_back(s.size());
  }
  return SetAttrStringList(attr_name,

  std::vector<int64_t> targets;
  std::vector<Gradient*> used_in_grads;
  // We may end up with slightly fewer elements than we reserve, but grad.size()
  // should be a reasonably tight upper bound.
  targets.reserve(grad.size());
  used_in_grads.reserve(grad.size());
  std::unordered_map<int64_t, TapeTensor> sources_that_are_targets;
  for (int grad_index = 0, end = grad.size(); grad_index < end; ++grad_index) {

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<std::unique_ptr<ParallelTensor>>& handles = *outputs;
  std::vector<ParallelTensor*> handle_inputs;
  handle_inputs.reserve(handles.size());
  for (auto& handle : handles) {
    handle_inputs.push_back(handle.get());
  }
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> read_op(

                                         tensorflow::Status* status) {
  std::vector<int64_t> shape;
  int rank = -1;
  *status = handle.NumDims(&rank);
  if (!status->ok()) {
    return shape;
  }
  shape.reserve(rank);
  for (int i = 0; i < rank; ++i) {
    int64_t dim;
    *status = handle.Dim(i, &dim);
    if (!status->ok()) {
      return shape;
    }
    shape.push_back(dim);
  }
  return shape;
}

    absl::flat_hash_set<int> null_indices;
    {
      AbstractContextPtr func_ctx(BuildFunction(fn_name));
      std::vector<AbstractTensorHandle*> func_inputs;
      func_inputs.reserve(inputs.size());
      TF_RETURN_IF_ERROR(
          CreateParamsForInputs(func_ctx.get(), inputs, &func_inputs));
      std::vector<AbstractTensorHandle*> model_outputs;
      model_outputs.resize(outputs.size());

        op_->colocation_constraints.emplace(static_cast<const char*>(values[i]),
                                            lengths[i]);
      }
    } else {
      std::vector<tensorflow::StringPiece> v;
      v.reserve(num_values);
      for (int i = 0; i < num_values; ++i) {
        v.emplace_back(static_cast<const char*>(values[i]), lengths[i]);
      }
      op_->node_builder.Attr(attr_name, v);
    }

  std::vector<TensorHandlePtr> release_tensors() { return std::move(tensors_); }

  std::vector<TFE_TensorHandle*> tensors() const {
    std::vector<TFE_TensorHandle*> result;
    result.reserve(tensors_.size());
    for (const TensorHandlePtr& tensor : tensors_) {
      result.emplace_back(tensor.get());
    }
    return result;
  }

 private:
  ParallelTensor(const ParallelDevice& device,