int64_t out_row_dim = output_shape[output_shape.size() - 2];
  int64_t out_col_dim = output_shape[output_shape.size() - 1];

  int64_t expected_out_row_dim = op.getAdjX() ? x_col_dim : x_row_dim;
  int64_t expected_out_col_dim = op.getAdjY() ? y_row_dim : y_col_dim;

  if (expected_out_row_dim != ShapedType::kDynamic &&
      out_row_dim != ShapedType::kDynamic &&
      out_row_dim != expected_out_row_dim)

    Variadic<TF_Tensor>:$output
  );

  TF_DerivedOperandTypeListAttr Tin = TF_DerivedOperandTypeListAttr<1>;
  TF_DerivedResultTypeListAttr Tout = TF_DerivedResultTypeListAttr<0>;
  TF_DerivedResultShapeListAttr output_shapes = TF_DerivedResultShapeListAttr<0>;

  let hasCanonicalizer = 1;

  let hasVerifier = 1;


 let extraClassDeclaration = [{
    int num_branches() { return getBranches().size(); }

  return %arg0 : tensor<!tf_type.variant>
}

// CHECK-LABEL: iterators
func.func private @iterators(%arg0 : tensor<!tf_type.variant>) {
  %0 = "tf.Iterator"() {container = "", output_shapes = [#tf_type.shape<200x28x28x1>, #tf_type.shape<200x10>], output_types = [f32, f32], shared_name = "_iterator1"} : () -> tensor<!tf_type.resource>
  %1 = func.call @id(%arg0) : (tensor<!tf_type.variant>) -> tensor<!tf_type.variant>

  // n_input_entries/n_output_entries
  // = group_size
  auto input_shape = Shape(scope, op.input(0));
  auto output_shape = Shape(scope, op.output(0));
  auto zero = Const(scope, 0);
  auto group_size = SafeDivHelper(scope, Prod(scope, input_shape, zero),
                                  Prod(scope, output_shape, zero));

  // propagate sum_grad/group_size
  grad_outputs->push_back(

"tf.While"(%cst_0, %cst, %cst_0, %arg0, %cst_1) {T = [i32, i32, i32, f32, f32],_lower_using_switch_merge = true, _num_original_outputs = 5 : i64, _read_only_resource_inputs = [], body = @while_body, cond = @while_cond, device = "", is_stateless = true, output_shapes = [#tf_type.shape<>, #tf_type.shape<>, #tf_type.shape<>, #tf_type.shape<1x1024>, #tf_type.shape<1024x1024>], parallel_iterations = 10 : i64, shape_invariant} : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<1x1024xf32>, tensor<1024x1024xf32>) ->...

    const xla::HloInputOutputAliasConfig& input_output_alias,
    absl::Span<const int> input_mapping,
    const std::map<int, const Tensor*>& resource_vars_snapshots,
    DataType output_dtype, const TensorShape& output_shape,
    Allocator* output_allocator, bool allocate_xla_tensors, se::Stream* stream,
    bool use_multiple_streams, std::shared_ptr<se::Event> definition_event) {

  return ConvertTensor(t, builder);
}

void ConvertToTensorShapeProto(ArrayRef<int64_t> shape,
                               TensorShapeProto* output_shape) {
  for (auto d : shape) {
    output_shape->add_dim()->set_size(ShapedType::isDynamic(d) ? kTFDynamicSize
                                                               : d);
  }
}

  %output_shape = arith.constant dense<[2, 3, 2, 2048]> : tensor<4xi32>
  %f16_weights = "tfl.pseudo_const"() {value = dense<1.0> : tensor<4x2x2x2048xf16>} : () -> tensor<4x2x2x2048xf16>
  %dq_weights = "tfl.dequantize"(%f16_weights) : (tensor<4x2x2x2048xf16>) -> tensor<4x2x2x2048xf32>
  %bias = "tfl.no_value"() {value} : () -> none

        init_func = @__init_func_530, next_func = @__next_func_680,
        next_func.experimental_ints_on_device = true,
        operandSegmentSizes = array<i32: 2, 2, 1>,
        output_shapes = [#tf_type.shape<?>],
        output_types = [f32],
        metadata = ""
      } :

        init_func = @__init_func_530, next_func = @__next_func_680,
        next_func.experimental_ints_on_device = true,
        operandSegmentSizes = array<i32: 2, 2, 1>,
        output_shapes = [#tf_type.shape<?>],
        output_types = [f32],
        metadata = ""
      } :