%6 = "tfl.quantize"(%5) {qtype = tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>, volatile} : (tensor<1x1x1x3xf32>) -> tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>
  %7 = "tfl.dequantize"(%6) : (tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>) -> tensor<1x1x1x3xf32>
  func.return %7 : tensor<1x1x1x3xf32>
// MODEL-DEBUG: %[[q1:.*]] = "tfl.quantize"(%arg0)

  %mm = "tfl.batch_matmul"(%w, %0) {adj_x = false, adj_y = true} : (tensor<1x1x12x512xf32>, tensor<1x1x3x512xf32>) -> tensor<1x1x12x3xf32>
  %mm_s = "quantfork.stats"(%mm) {layerStats = dense<[0.000000e+00, 1.000000e+01]> : tensor<2xf32>} : (tensor<1x1x12x3xf32>) -> tensor<1x1x12x3xf32>
  func.return %mm_s : tensor<1x1x12x3xf32>

// CHECK: %[[w:.*]] = arith.constant dense<1.270000e+02> : tensor<1x1x12x512xf32>

func.func private @conv(%input: tensor<1x3x4x3xf32> {tf._user_specified_name = "input_tensor"}) -> tensor<*xf32> attributes {tf._construction_context = "kEagerRuntime", tf._input_shapes = [#tf_type.shape<1x3x4x3>]} {
  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>
  %bias = arith.constant dense<[7.11401462, 7.05456924]> : tensor<2xf32>

  } : (tensor<1x3x4x3x3xf32>, tensor<2x3x3x3x2xf32>) -> tensor<1x3x2x3x2xf32>
  %7 = "tf.BiasAdd"(%6, %cst_1) {data_format = "NHWC", device = ""} : (tensor<1x3x2x3x2xf32>, tensor<2xf32>) -> tensor<1x3x2x3x2xf32>

  return %2, %5, %7 : tensor<1x3x2x3x2xf32>, tensor<1x3x2x3x2xf32>, tensor<1x3x2x3x2xf32>

// CHECK-DAG: %[[CST:.*]] = "tf.Const"() {{.*}} : () -> tensor<2x3x3x3x2xf32>

func.func @lift_float_conv3d(%arg0: tensor<1x3x4x3x3xf32>) -> (tensor<1x3x2x3x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.0> : tensor<2x3x3x3x2xf32>} : () -> tensor<2x3x3x3x2xf32>
  %0 = "tf.Conv3D"(%arg0, %cst) {
    data_format = "NDHWC", device = "", dilations = [1, 1, 1, 1, 1], padding = "SAME", strides = [1, 1, 2, 1, 1]
  } : (tensor<1x3x4x3x3xf32>, tensor<2x3x3x3x2xf32>) -> tensor<1x3x2x3x2xf32>

// RUN: tf-quant-opt %s -split-input-file -quant-insert-quantized-functions -quant-quantize-composite-functions | FileCheck %s

module {
  func.func @conv(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {

func.func @add_with_activation_transpose(%arg0: tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<1x4x3x3xf32>
  %1 = stablehlo.transpose %arg0, dims = [0, 3, 1, 2] : (tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>
  %2 = stablehlo.add %1, %0 : tensor<1x4x3x3xf32>
  return %2 : tensor<1x4x3x3xf32>
}
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>

  %5 = "tfl.quantize"(%4) {qtype = tensor<1x34x1x3x!quant.uniform<i8:f32, 1.0>>, volatile} : (tensor<1x34x1x3xf32>) -> tensor<1x34x1x3x!quant.uniform<i8:f32, 1.0>>
  %6 = "tfl.dequantize"(%5) : (tensor<1x34x1x3x!quant.uniform<i8:f32, 1.0>>) -> tensor<1x34x1x3xf32>

  %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>
}

    %1 = "tf.PartitionedCall"(%arg0, %cst, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_conv2d_with_bias_and_relu6_fn_1} : (tensor<1x2x2x3xf32>, tensor<2x2x3x2xf32>, tensor<2xf32>) -> tensor<*xf32> loc(callsit...