%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
    %14 = call @uniform_quantize_1(%13, %5, %6) : (tensor<1x3xf32>, tensor<1x1xf32>, tensor<1x1xi8>) -> tensor<1x3xi8>

    %9 = stablehlo.select %arg2, %6, %8 : (tensor<1x3xi1>, tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
    %10 = "quantfork.qcast"(%9) {volatile} : (tensor<1x3xf32>) -> tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
    %11 = "quantfork.dcast"(%10) : (tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<1x3xf32>
    return %11 : tensor<1x3xf32>
  }

      %cst = stablehlo.constant dense<1.0> : tensor<1x3xf32>
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      %1 = stablehlo.subtract %cst, %0 : tensor<1x3xf32>
      %2 = stablehlo.add %0, %cst : tensor<1x3xf32>
      return %2 : tensor<1x3xf32>
    }
  }
)mlir";