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>

    return %1 : tensor<1x3x4x2xf32>
  }

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

    %12 = "quantfork.qcast"(%11) {volatile} : (tensor<2x3x1x3xf32>) -> tensor<2x3x1x3x!quant.uniform<i8:f32, 3.000000e-01:1>>
    %13 = "quantfork.dcast"(%12) : (tensor<2x3x1x3x!quant.uniform<i8:f32, 3.000000e-01:1>>) -> tensor<2x3x1x3xf32>
    return %13 : tensor<2x3x1x3xf32>
  }

  // CHECK: quantized_dot_general_fn_1

    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

      stablehlo.return %2 : tensor<f32>
  }) {padding = dense<[[0, 0], [1, 1], [1, 1], [0, 0]]> : tensor<4x2xi64>, window_dimensions = array<i64: 1, 3, 3, 1>} : (tensor<2x3x1x3xf32>, tensor<f32>) -> tensor<2x3x1x3xf32>
  return %1 : tensor<2x3x1x3xf32>
}

// -----

// CHECK-LABEL: @bitcast_convert_i32_f32(%arg0: tensor<1x256128xi32>) -> tensor<1x256128xbf16>

  return %4 : tensor<1x3x2x2xf32>
}

// CHECK: func @cast_bf16_conv_with_bias_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>
// CHECK-DAG: %[[cst_0:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<2xf32>}> : () -> tensor<2xf32>
// CHECK: %[[conv:.*]] = "tf.Conv2D"(%arg0, %[[cst]])

    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

    %cst_0 = "tf.Const"() {device = "", value = dense<[[[[0.211401448, 0.205456927], [0.418355644, -0.314615548]], [[0.493921608, -0.101286061], [-0.16083248, -0.0546654463]], [[-0.157245964, 0.419805884], [-0.0499645844, 3.726310e-01]]],...

  %0 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<16x3x3x3xf32>} : () -> tensor<16x3x3x3xf32>
  %1 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<16xf32>} : () -> tensor<16xf32>
  %2 = func.call @fold_all_test(%arg0, %0, %1) : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>
  func.return %2 : tensor<256x30x30x16xf32>
}

    %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>
    %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}