%9 = stablehlo.convert %2 : (tensor<2x3xi32>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %12 = stablehlo.subtract %10, %11 : tensor<1x3xf32>  // q1 * q2 - z1 * q2
    %13 = stablehlo.multiply %12, %4 : tensor<1x3xf32>  // s1 * s2

  func.func private @composite_and_select(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>, %arg2: tensor<1x3xi1>, %arg3: tensor<1x3xf32>) -> tensor<1x3xf32> {
    // CHECK: %[[Q1:.*]] = "quantfork.qcast"(%[[ARG0]]) {volatile} : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<i8:f32, 5.000000e-03>>

}

// CHECK-LABEL: QuantizeConcat
func.func @QuantizeConcat(tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.000000e-01:128>> {
^bb0(%arg0: tensor<1x2xf32>, %arg1: tensor<1x2xf32>):
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2xf32>

      %2 = "tf.XlaCallModule"(%arg0, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_original_entry_function = "composite_fn_1"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>
      return %2 : tensor<?x2xf32>
    }

  %12 = mhlo.clamp %9, %11, %1 : tensor<2xf32>
  %13 = mhlo.multiply %arg0, %12 : tensor<2xf32>
  %14 = mhlo.divide %13, %1 : tensor<2xf32>
  return %14 : tensor<2xf32>
}

// CHECK-LABEL:   func.func @hardswish(
// CHECK-SAME:                         %[[VAL_0:.*]]: tensor<2xf32>) -> tensor<*xf32> {
// CHECK:           %[[VAL_1:.*]] = "tfl.hard_swish"(%[[VAL_0]]) : (tensor<2xf32>) -> tensor<2xf32>