%0 = "tfl.pseudo_qconst"() {qtype = tensor<1x!quant.uniform<i8:f32, 0.003:-128>>, value = dense<127> : tensor<1xi8>} : () -> tensor<1x!quant.uniform<i8:f32, 0.003:-128>>
  %1 = "tfl.pseudo_qconst"() {qtype = tensor<1x!quant.uniform<i8:f32, 0.003:-128>>, value = dense<127> : tensor<1xi8>} : () -> tensor<1x!quant.uniform<i8:f32, 0.003:-128>>

    %19 = "tf.BroadcastTo"(%4, %17) : (tensor<2x?x?xi8>, tensor<3xi64>) -> tensor<2x?x?xi8>
    %20 = "tf.BroadcastTo"(%9, %18) : (tensor<2x?x?xi8>, tensor<3xi64>) -> tensor<2x?x?xi8>
    %21 = "tf.XlaDotV2"(%19, %20) {dimension_numbers = "\22\01\00\1A\01\00\12\01\01\0A\01\02", precision_config = ""} : (tensor<2x?x?xi8>, tensor<2x?x?xi8>) -> tensor<2x?x?xi32>
    %22 = "tf.Cast"(%19) {Truncate = false} : (tensor<2x?x?xi8>) -> tensor<2x?x?xi32>

          )
      )
      module_str = self._extract_first_xla_call_module_op(
          self._output_saved_model_path
      )
      self.assertTrue(
          re.search('stablehlo.dot_general.*xi8>.*xi8>.*xi32>', module_str)
      )

    # Run quantization the first time, calibration is expected to be skipped.
    output_saved_model_path_2 = self.create_tempdir('output2').full_path