->DoHostCallback([device_to_host_stream] {})
                .IgnoreError();
          }
        }
      });
}

se::Stream* XlaDeviceContext::GetDeviceToDeviceStream() {
  DCHECK_GT(device_to_device_streams_.size(), 0);
  absl::MutexLock lock(&mu_);
  int stream = next_stream_;
  next_stream_ = (next_stream_ + 1) % device_to_device_streams_.size();
  return device_to_device_stream(stream);
}

  ~CApiWhileLoopTest() override {
    TF_DeleteGraph(graph_);
    TF_DeleteStatus(s_);
  }

  void Init(int ninputs) {
    DCHECK(inputs_.empty());
    DCHECK_GT(ninputs, 0);

    for (int i = 0; i < ninputs; ++i) {
      TF_Operation* placeholder = Placeholder(
          graph_, s_, ::tensorflow::strings::StrCat("p", i).c_str());

  // processed yet. This happens if not all outputs of a node are in 'inputs_'.
  std::unordered_map<Node*, int> requested_grads;
  for (const Output& nout : inputs_) {
    if (pending_[nout.node()->id()] > 0) {
      DCHECK_GT(nout.node()->num_outputs(), 1);
      int idx = input_nodes_[nout];
      DCHECK(((*grad_outputs_)[idx].node() == nullptr));
      TF_RETURN_IF_ERROR(SumGradients(nout, &(*grad_outputs_)[idx]));