%cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
    %cst_2 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<*xf32>

    %0 = "tf.DepthwiseConv2dNative"(%arg0, %cst_0) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1]} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<1x2x2x3xf32>

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
    %cst_2 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<*xf32>

func.func @QuantizeGatherWeightOnly(%arg0: tensor<3xi32>) -> tensor<3x3x3x3xf32> {
  %w = arith.constant dense<1.270000e+02> : tensor<64x3x3x3xf32>
  %emb = "tfl.gather"(%w, %arg0) {axis = 0 : i32, batch_dims = 0 : i32} : (tensor<64x3x3x3xf32>, tensor<3xi32>) -> tensor<3x3x3x3xf32>
  %emb_s = "quantfork.stats"(%emb) {layerStats = dense<[0.000000e+00, 1.000000e+01]> : tensor<2xf32>} : (tensor<3x3x3x3xf32>) -> tensor<3x3x3x3xf32>
  func.return %emb_s : tensor<3x3x3x3xf32>

  %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
  %0 = "tf.Conv2D"(%arg0, %cst_1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true
  } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

  %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>

  %cst_1 = arith.constant dense<6.000000e+00> : tensor<2x1x3x3xf32>
  %0 = "tfl.quantize"(%cst_1) {qtype = tensor<2x1x3x3x!quant.uniform<i8<-127:127>:f32:0, {6.587140e-03,1.888450e-02}>>} : (tensor<2x1x3x3xf32>) -> tensor<2x1x3x3x!quant.uniform<i8<-127:127>:f32:0, {6.587140e-03,1.888450e-02}>>
  %1 = "tfl.dequantize"(%0) : (tensor<2x1x3x3x!quant.uniform<i8<-127:127>:f32:0, {6.587140e-03,1.888450e-02}>>) -> tensor<2x1x3x3xf32>

  func.return %0 : tensor<256x32x32x16xf32>
}

// -----

func.func @testConv2D4DBias(tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<1x1x1x16xf32>) -> tensor<256x32x32x16xf32> {
^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<1x1x1x16xf32>):

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 4, narrow_range = true} : (tensor<3x3x3x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<3x3x3x4xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x4xf32>) -> tensor<256x8x7x4xf32>
  func.return %rst : tensor<256x8x7x4xf32>

  %cst_0 = arith.constant dense<[1.0, 2.0, 3.0]> : tensor<3xf32>
  %w = arith.constant dense<2.0> : tensor<3x3x3x3xf32>
  %q = "tfl.quantize"(%w) {qtype = tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>} : (tensor<3x3x3x3xf32>) -> tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>
  %dq = "tfl.dequantize"(%q) : (tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>) -> tensor<3x3x3x3xf32>