mlir::tf_device::ClusterFuncOp cluster_func, mlir::OpBuilder* builder,
    llvm::SmallVectorImpl<llvm::SmallVector<mlir::Value, 4>>* input_list) {
  // Initialize the input list for each logical devices.
  input_list->reserve(num_cores_per_replica);
  for (int i = 0; i < num_cores_per_replica; ++i)
    input_list->emplace_back(llvm::SmallVector<mlir::Value, 4>());

  llvm::SmallVector<mlir::Value, 4> cluster_func_inputs(

  // i-th logical core.
  llvm::SmallVector<llvm::SmallVector<mlir::Value, 4>, 4> input_list;
  builder->setInsertionPoint(*new_parallel_execute);
  auto result = tensorflow::ExtractInputsForLogicalDevices(
      num_cores_per_replica, cluster_func, builder, &input_list);
  if (failed(result)) return failure();

  const bool replicated = tpu_devices.size() != 1;

                                  next_output_index, dtype);

  std::vector<NodeBuilder::NodeOut> input_list({enter_input, next_input});
  const string unique_name = scope.GetUniqueNameForOp("Merge");
  NodeBuilder builder = NodeBuilder(unique_name, "Merge").Input(input_list);
  scope.UpdateBuilder(&builder);

  Node* merge_node;
  TF_RETURN_IF_ERROR(builder.Finalize(scope.graph(), &merge_node));

mlir::LogicalResult ExtractInputsForLogicalDevices(
    int num_cores_per_replica, mlir::tf_device::ClusterFuncOp cluster_func,
    mlir::OpBuilder* builder,
    llvm::SmallVectorImpl<llvm::SmallVector<mlir::Value, 4>>* input_list);

// Extracts a list of OpSharding that represent output sharding configuration of
// `tf_device.cluster`.
mlir::LogicalResult ParseAndValidateOutputSharding(

}

// Creates a slice of the tensorlist `input_list`, starting from
// [start_index, 0, ...0], with size [size, -1, ...-1].
//
// Requires that `start_index` and `size` are scalar tensors and
// `item_position_shape` is a 1-D tensor with only one element equal to the rank
// of an item in the tensorlist.
TF::SliceOp CreateSliceOpForTensorList(Location loc, Value input_list,

                     int num_inputs) {
  std::vector<NodeBuilder::NodeOut> input_list;
  input_list.reserve(num_inputs);
  for (int i = 0; i < num_inputs; ++i) {
    input_list.emplace_back(&inputs[i].oper->node, inputs[i].index);
  }
  desc->node_builder.Input(input_list);
}

void TF_AddControlInput(TF_OperationDescription* desc, TF_Operation* input) {

class InputList {
 public:
  /// Implicitly convert a list of outputs to a list of inputs. This is useful
  /// to write code such as ops::Concat(ops::Split(x, 4)).
  InputList(const OutputList& out) {  // NOLINT(runtime/explicit)
    for (auto const& x : out) {
      inputs_.push_back(x);
    }
  }

  InputList(

    final ImmutableList<E> inputList;
    final Comparator<? super E> comparator;
    final int size;

    OrderedPermutationCollection(Iterable<E> input, Comparator<? super E> comparator) {
      this.inputList = ImmutableList.sortedCopyOf(comparator, input);
      this.comparator = comparator;
      this.size = calculateSize(inputList, comparator);
    }

    /**

    final ImmutableList<E> inputList;
    final Comparator<? super E> comparator;
    final int size;

    OrderedPermutationCollection(Iterable<E> input, Comparator<? super E> comparator) {
      this.inputList = ImmutableList.sortedCopyOf(comparator, input);
      this.comparator = comparator;
      this.size = calculateSize(inputList, comparator);
    }

    /**

  // CHECK: %[[input:.*]] = "tf.ConcatV2"(%arg0, %arg1, %[[VAL:.*]]) : (tensor<3x3xi8>, tensor<3x3xi8>, tensor<i64>) -> tensor<6x3xi8>
  // CHECK: %[[input_qint:.*]] = "tf.Cast"(%[[input]]) <{Truncate = false}> : (tensor<6x3xi8>) -> tensor<6x3x!tf_type.qint8>
  // CHECK: %[[output:.*]] = "tf.UniformDequantize"(%[[input_qint]]
  // CHECK: return %[[output]] : tensor<6x3xf32>