%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>

    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> 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 @gelu_decomp_2(%arg0: tensor<5x10xf32>) -> tensor<5x10xf32>
func.func @gelu_aten_approximate(%arg0: tensor<5x10xf32>) -> (tensor<*xf32>) {
  %0 = mhlo.composite "aten.gelu.default" %arg0 {composite_attributes = {approximate = "tanh"}, decomposition = @gelu_decomp_2} : (tensor<5x10xf32>) -> tensor<5x10xf32>

// Test for XlaLaunch

func.func private @xla_func_0(%arg0: tensor<1x3xf32>, %arg1: tensor<1x3xf32>) -> tensor<1x3xf32> attributes {tf._XlaMustCompile = true, tf._noinline = true, tf._original_func_name = "should_not_be_used"} {
  %1 = "tf.AddV2"(%arg0, %arg1) {__op_key = 0: i32} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %1 : tensor<1x3xf32>
}

// CHECK-LABEL: func @xla_func