%2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>
      return %2 : tensor<1x3xf32>
    }

    func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {

    %5 = "quantfork.qcast"(%4) {volatile} : (tensor<1x3xf32>) -> tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
    %6 = "quantfork.dcast"(%5) : (tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<1x3xf32>
    %7 = stablehlo.reshape %6 : (tensor<1x3xf32>) -> tensor<3x1xf32>

      time_major = false} : (
        tensor<1x2x3xf32>,
        tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>,
        tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>,
        none, none, none,
        tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>,
        none, none,
        tensor<1x3xf32>, tensor<1x3xf32>,
        none, none, none, none) -> tensor<1x2x3xf32>

      input_to_forget_intermediate = tensor<0xf32>,
      input_to_input_intermediate = tensor<0xf32>,
      input_to_output_intermediate = tensor<0xf32>,
      proj_clip = 0.000000e+00 : f32,
      time_major = false} : (
        tensor<1x2x3xf32>,
        tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>,
        tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>,

    return %1 : tensor<1x3xf32>
  }
}

// -----

// Do not merge when function is not quantized function.

module attributes {tf_saved_model.semantics} {
  // CHECK-LABEL: func.func private @merge_relu_fusion
  func.func private @merge_relu_fusion(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {

  %outputs_2 = "tf.add"(%outputs_0, %outputs_1): (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x1xf32>
  return %outputs_2 : tensor<1x1xf32>
}

// CHECK-LABEL: @_func
func.func private @_func(%arg0: tensor<1x3xf32>, %arg1: tensor<3x1xf32>) -> (tensor<1x1xf32>) {

  // CHECK-SAME: %[[ARG3:[^:[:space:]]+]]
  func.func private @main(%arg0: tensor<1x4xf32>, %arg1: tensor<4x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>
    // CHECK: %[[DOT:.+]] = stablehlo.dot_general %[[ARG2]], %[[ARG3]]
    // CHECK: return %[[DOT]]
  }

// CHECK-NEXT:    "tf.MatMul"(%arg0, [[TENSOR]]) : (tensor<1x3xf32>, tensor<3x1xf32>) -> tensor<1x1xf32>
// CHECK-NEXT:    "tf.IfrtCall"(%arg0, [[ARRAYKEY]]) <{program_id = 6515870160938153680 : i64, variable_arg_indices = [1 : i32]}> {__tpu_compile_metadata_text = "retvals { sharding { } }"} : (tensor<1x3xf32>, tensor<!tf_type.string>) -> tensor<1x1xf32>
// CHECK-NEXT:    return
//

  %add = stablehlo.add %dq1, %dq2 : (tensor<3x2xf32>, tensor<3x2xf32>) -> tensor<3x2xf32>
  func.return %add : tensor<3x2xf32>
}

// -----

// CHECK-LABEL: @quantize_constant
// CHECK-SAME: %[[ARG0:.*]]: tensor<1x3xf32>
func.func @quantize_constant(%arg0: tensor<1x3xf32>) -> tensor<1x2xf32> {

    return %3 : tensor<1x3xf32>
  }
  func.func private @composite_matmul_with_bias_fn_2(%arg0: tensor<1x4xf32>, %arg1: tensor<4x3xf32>, %arg2: tensor<3xf32>) -> tensor<1x3xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.MatMul"(%arg0, %arg1) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> {attr_map = "0:transpose_a,1:transpose_b", device = ""} : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>