// allocation/scheduling.
  //
  // TODO(allenl): Keep a map from outer thread to list of inner threads rather
  // than a single list of threads so aliased nested parallel devices don't
  // 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

    return absl::OkStatus();
  }
  if (!ShouldRecord(input_tensor_id, input_dtypes)) {
    return absl::OkStatus();
  }

  // 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) {