}

  Status AddInput(AbstractTensorHandle* input) override {
    GraphTensor* t = dyn_cast<GraphTensor>(input);
    if (!t) {
      return tensorflow::errors::InvalidArgument(
          "Unable to cast input to GraphTensor");
    }
    TF_AddInput(op_.get(), t->output_);
    return absl::OkStatus();
  }
  Status AddInputList(absl::Span<AbstractTensorHandle* const> inputs) override {

// CHECK: %[[lstm:.*]] = "tfl.unidirectional_sequence_lstm"(%[[input_0]], %[[input_1]], %[[input_2]], %[[input_3]], %[[input_4]], %[[input_5]], %[[input_6]], %[[input_7]], %[[input_8]],
// CHECK-SAME: %[[input_9]], %[[input_10]], %[[input_11]], %[[input_12]], %[[input_13]], %[[input_14]], %[[input_15]], %[[input_16]], %[[input_17]], %[[input_18]], %[[input_19]]

// CHECK-SAME: %[[input_0]],
// CHECK-SAME: %[[input_1]], %[[input_2]], %[[input_3]], %[[input_4]],
// CHECK-SAME: %[[input_5]], %[[input_6]], %[[input_7]], %[[input_8]],
// CHECK-SAME: %[[input_9]], %[[input_9]], %[[input_9]],
// CHECK-SAME: %[[input_10]], %[[input_11]], %[[input_12]], %[[input_13]],
// CHECK-SAME: %[[input_9]], %[[input_9]],
// CHECK-SAME: %[[input_14]], %[[input_15]],

is used instead.

Operands are replicated inputs and packed inputs.

replicated_inputs: each group of `n` inputs corresponds to an input for a single
individual replica and is mapped to a single region argument. Inside one group
the operands are matching in order the `devices` attribute. Each replicated
input must have compatible shapes and types.
packed_inputs: each input corresponds to an input broadcasted across all

    RunOptions run_options;

    Tensor input_a = CreateInputTensor(shape, 0);
    Tensor input_b = CreateInputTensor(shape, shape.num_elements());
    Tensor input_c = CreateInputTensor(shape, 2 * shape.num_elements());
    TF_RETURN_IF_ERROR(session->Run(
        run_options,
        {std::make_pair("a", input_a), std::make_pair("b", input_b),
         std::make_pair("c", input_c)},
        {"m"}, {}, &golden_output_tensors, nullptr));

  std::vector<TFE_TensorHandle*> compute_fn_inputs;
  TFE_TensorHandle* input_a = TestScalarTensorHandle(ctx, 2.0f);
  TFE_TensorHandle* input_b = TestScalarTensorHandle(ctx, 1.0f);
  compute_fn_inputs.push_back(input_a);
  compute_fn_inputs.push_back(input_b);

  TFE_Op* compute_fn_op = TF_ConcreteFunctionMakeCallOp(

  // TFL lstm only supports time-majored inputs, so if it's not time-majored,
  // we will transpose the inputs and outputs.
  auto time_major_attr = func_op->getAttrOfType<BoolAttr>("tf.time_major");
  if (time_major_attr == nullptr) return failure();

  bool time_majored = time_major_attr.getValue();
  auto input_type = mlir::dyn_cast_or_null<RankedTensorType>(input.getType());
  if (!input_type) {

  func::FuncOp fused_func_op_;
  Value input_;
  Value weight_;
  Value bias_;
  Value projection_;
  bool couple_input_forget_gates_;

  // internal state
  Value weight_transposed_;
  Value projection_transposed_;
  RankedTensorType weight_type_;
  RankedTensorType projection_type_;
  int num_gates_;
  int n_cell_;
  int n_output_;
  int n_input_;
  int num_cols_weight_transposed_;

    }

    self._assertOpAndComposite([input_, filter_, bias],
                               tf.function(gen_mnist_ops.new_conv2d),
                               ops_defs._composite_conv_add_relu, kwargs)

  def test_new_conv2d_relu6(self):
    input_ = tf.random.uniform([1, 4, 4, 1])
    filter_ = tf.random.uniform([2, 2, 1, 8])
    bias = tf.zeros([8])
    kwargs = {
        'input_': input_,
        'filter_': filter_,

    Node* input =
        ops::SourceOp("FakeNullary", builder.opts().WithName("Input"));
    Node* clustered_producer =
        ops::BinaryOp("FakeBinary", input, input,
                      builder.opts().WithName("ClusteredProducer"));
    // The first input is hostmem and the second input is devicemem.
    Node* consumer_in_different_cluster =
        ops::BinaryOp("FakeBinary", input, clustered_producer,