components.reserve(underlying_devices_.size());
    for (int j = 0; j < underlying_devices_.size(); ++j) {
      components.push_back(std::move(per_device_output_tensors[j][i]));
    }
    if (expected_output_shapes[i].IsFullyDefined()) {
      per_device_outputs.push_back(ParallelTensor::FromTensorHandles(
          *this, std::move(components),
          absl::Span<const int64_t>(expected_output_shapes[i].dim_sizes()),

  const auto& shape = reader->GetVariableToShapeMap().at(name);
  int rank = shape.dims();
  if (num_dims != rank) {
    status->status = InvalidArgument("Expected rank is ", num_dims,
                                     " but actual rank is ", rank);
    return;
  }
  for (int i = 0; i < num_dims; i++) {
    dims[i] = shape.dim_size(i);
  }
}

        TFE_DeleteTensorHandle(outputs[i]);
      }
    }
    return status.status;
  }

  tensorflow::Status Pack(absl::Span<ImmediateExecutionTensorHandle*> handles,
                          ImmediateExecutionTensorHandle** result) override {
    TF_Status status;
    *result = tensorflow::unwrap(device_.pack(context_,
                                              tensorflow::wrap(handles.data()),

  public <T extends @Nullable Object> T[] toArray(T[] array) {
    return snapshot().toArray(array);
  }

  /*
   * We'd love to use 'new ArrayList(this)' or 'list.addAll(this)', but
   * either of these would recurse back to us again!
   */
  private List<E> snapshot() {
    List<E> list = Lists.newArrayListWithExpectedSize(size());
    for (Multiset.Entry<E> entry : entrySet()) {
      E element = entry.getElement();

     *       cleared, we know that we're not executing inline any more
     * </ul>
     *
     * All the states where thread != currentThread are identical for our purposes, and so even
     * though it's racy, we don't care which of those values we get, so no need to synchronize.
     */
    @CheckForNull @LazyInit Thread thread;

    /** Only used by the thread associated with this object */

  forward_op_->inputs.push_back(input);
  return absl::OkStatus();
}
Status AddInputList(AbstractOperation* op_,
                    absl::Span<AbstractTensorHandle* const> inputs,
                    ForwardOperation* forward_op_) {
  TF_RETURN_IF_ERROR(op_->AddInputList(inputs));
  for (auto input : inputs) {
    forward_op_->inputs.push_back(input);
  }
  return absl::OkStatus();
}

   *
   * 1. visibility of the writes to these fields to Fire.run:
   *
   * The initial write to delegateRef is made definitely visible via the semantics of
   * addListener/SES.schedule. The later racy write in cancel() is not guaranteed to be observed,
   * however that is fine since the correctness is based on the atomic state in our base class. The

    std::vector<PartialTensorShape> shapes;
    shapes.reserve(num_values);
    for (int i = 0; i < num_values; ++i) {
      if (num_dims[i] < 0) {
        shapes.emplace_back();
      } else {
        shapes.emplace_back(ArraySlice<int64_t>(
            reinterpret_cast<const int64_t*>(dims[i]), num_dims[i]));
      }
    }
    op_->node_builder.Attr(attr_name, shapes);
    return absl::OkStatus();
  }

This property is sometimes referred to as DRF-SC: data-race-free programs
execute in a sequentially consistent manner.
</p>

<p>
While programmers should write Go programs without data races,
there are limitations to what a Go implementation can do in response to a data race.
An implementation may always react to a data race by reporting the race and terminating the program.

  public <T extends @Nullable Object> T[] toArray(T[] array) {
    return snapshot().toArray(array);
  }

  /*
   * We'd love to use 'new ArrayList(this)' or 'list.addAll(this)', but
   * either of these would recurse back to us again!
   */
  private List<E> snapshot() {
    List<E> list = Lists.newArrayListWithExpectedSize(size());
    for (Multiset.Entry<E> entry : entrySet()) {
      E element = entry.getElement();