%4 = "quantfork.stats"(%3) {
      layerStats = dense<[-1.0, 0.8]> : tensor<2xf32>
    } : (tensor<*xf32>) -> tensor<*xf32>
    %5 = "tf.PartitionedCall"(%4, %cst_1, %cst_2) {
      _tfl_quant_trait = "fully_quantizable", config = "", config_proto = "",
      executor_type = "", f = @composite_matmul_with_bias_fn_1

func.func @serving_default(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x2x2x2xf32>) attributes {tf.entry_function = {control_outputs = "", inputs = "serving_default_input_tensor:0", outputs = "PartitionedCall:0"}} {

      attributes {tf_quant.quantized_ops = ${quantized_ops}} {

    %input_scale, %input_zp = "tf.PartitionedCall"(%input) {
        config = "", config_proto = "", executor_type = "", f=@internal_calculate_quant_params
      } : (tensor<*xf32>) -> (tensor<*xf32>, tensor<*xi32>)

    %quantized_input = "tf.PartitionedCall"(%input, %input_scale, %input_zp) {
        config = "", config_proto = "", executor_type = "", f=@internal_quantize_i8

    %0 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 149.824783]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

      attributes {tf_quant.quantized_ops = ${quantized_ops}} {

    %accum_out = "tf.PartitionedCall"(%input, %weight) {
        config = "", config_proto = "", executor_type = "", f=@${internal_func_name}
      } : (tensor<*xf32>, tensor<*xi8>) -> tensor<*xf32>

    %out = "tf.PartitionedCall"(%accum_out, %weight_scale) {
        config = "", config_proto = "", executor_type = "", f=@internal_dequantize_f32

// CHECK-NEXT:    %1 = "tf.AddV2"(%0, %cst) : (tensor<i32>, tensor<i32>) -> tensor<i32>
// CHECK-NEXT:    "tf_mlrt.tf_promise"(%arg2, %1) : (!mlrt.promise, tensor<i32>) -> ()
// CHECK-NEXT:    %2 = "tf.PartitionedCall"(%1, %arg5) <{config = "", config_proto = "", executor_type = "", f = @"map/while_cond/TfMlrtAsyncWhilePredicate"}> : (tensor<i32>, tensor<i32>) -> tensor<i1>

    %cst = "tf.Const"() {device = "", value = dense<[[[[-0.315365672, 0.27481091], [0.0901821703, -0.382271349], [-0.105572946, -0.354302853]], [[-0.47703138, -0.307006568],...

module attributes {tf.versions = {bad_consumers = [], min_consumer = 12 : i32, producer = 29 : i32}} {
  func.func @main() {
    tf_executor.graph {
      %0 = tf_executor.island wraps "tf.PartitionedCall"() {Tin = [], Tout = [], config = "", config_proto = "", device = "", executor_type = "", f = @foo, name = "Call_foo"} : () -> ()
      tf_executor.fetch
    }
    func.return
  }
  func.func @foo() {

    %1 = "tf.Reshape"(%arg1, %0) : (tensor<19x10xf32>, tensor<2xi64>) -> tensor<?x?xf32>

    // CHECK: %[[RESULT:.*]] = mhlo.constant dense<[10, 19]>
    %2 = "tf.PartitionedCall"(%1) {config = "", config_proto = "", executor_type = "", f = @get_shape} : (tensor<?x?xf32>) -> (tensor<?xi64>)

    // CHECK: return %[[RESULT]]
    func.return %2 : tensor<?xi64>
  }

  // CHECK: _xla_original_oc_node_name = [[NODE_NAME1:.*]], _xla_token_input_nodes = ["_xla_token_arg_node"]
  %0 = "tf.PartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @PartitionedCall3} : (tensor<i32>) -> (tensor<i32>)
  // CHECK-NOT: _xla_token_input_nodes
  %1 = "tf.PartitionedCall"(%0) {config = "", config_proto = "", executor_type = "", f = @IdentityFunc} : (tensor<i32>) -> (tensor<i32>)