} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>
    return %0 : tensor<1x3x4x2xf32>
  }

  // CHECK: func.func private @qdq_for_conv_weight_per_channel(%[[ARG0:.+]]: tensor<1x3x4x3xf32>)
  // CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

  %0 = stablehlo.constant dense<2.000000e+00> : tensor<1x3x2x4xf32>
  %1 = stablehlo.transpose %arg0, dims = [0, 2, 1, 3] : (tensor<1x2x3x4xf32>) -> tensor<1x3x2x4xf32>
  %2 = stablehlo.add %1, %0 : tensor<1x3x2x4xf32>
  return %2 : tensor<1x3x2x4xf32>
}
// CHECK: %[[TRANSPOSE_0:.+]] = stablehlo.transpose
// CHECK: %[[ADD_0:.+]] = stablehlo.add %[[TRANSPOSE_0]], {{.*}}
// CHECK: return %[[ADD_0]]

// -----

        func.return %2: tensor<1x3x3x4xf32>
      }

      func.func @conv_1_fn(%arg0: tensor<1x3x3x4xf32>) -> tensor<1x3x3x4xf32> {
        %0 = stablehlo.constant dense<2.000000e+00> : tensor<3x3x4x4xf32>

    %16 = stablehlo.subtract %10, %15 : tensor<1x3x3x4xf32>
    %17 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2, 3] : (tensor<1x1x1x4xf32>) -> tensor<1x3x3x4xf32>
    %18 = stablehlo.multiply %16, %17 : tensor<1x3x3x4xf32>

  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", device = "", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

    %2 = "quantfork.stats"(%1) {layerStats = dense<[4.00000000e-6, 9.80000000e-1]> : tensor<2xf32>} : (tensor<2x3x2x2xf32>) -> tensor<2x3x2x2xf32>
    return %2 : tensor<2x3x2x2xf32>
  }
// Checks that the quantized XlaCallModule has been replaced by a CallOp, which
// calls the quantized entry function.

  %cst_1 = "tf.Const"() {value = dense<[0.1, 0.2]> : tensor<2xf32>} : () -> tensor<2xf32>
  %cst_2 = "tf.Const"() {value = dense<[0.3, 0.4]> : tensor<2xf32>} : () -> tensor<2xf32>

  func.func private @composite_conv2d_with_bias_and_relu6_fn_10(%arg0: tensor<1x3x2x3xf32>, %arg1: tensor<2x3x3x2xf32>, %arg2: tensor<2xf32>) -> tensor<1x2x2x2xf32> attributes {tf.tf_quant.composite_function} {
    %0 = "quantfork.stats"(%arg1) {layerStats = dense<[-3.54062747, 0.54742622]> : tensor<2xf32>} : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2xf32>

        %cst = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
        %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
        %cst_1 = "tf.Const"() {value = dense<0.500000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
        %w = "tf.Mul"(%cst, %arg1) : (tensor<2x3x3x1xf32>, tensor<f32>) -> tensor<2x3x3x1xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {