%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>

  %w = arith.constant dense<127.0> : tensor<3x3x3x3xf32>
  %b = arith.constant dense<0.0> : tensor<3xf32>
  %conv = "tfl.conv_2d"(%arg0, %w, %b) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<1x5x5x3xf32>, tensor<3x3x3x3xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  func.return %conv : tensor<1x5x5x3xf32>

  %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>

        %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>

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>

  %1 = "tf.Identity"(%0) {device = ""} : (tensor<1x3x4x4xf32>) -> tensor<*xf32>
  %2 = "tf.Identity"(%1) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>
  return %2 : tensor<*xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_0(%arg0: tensor<1x3x6x6xf32>) -> tensor<1x3x4x4xf32> {
  %0 = mhlo.constant dense<1.000000e+00> : tensor<f32>

  %q_weight = "quantfork.qcast"(%weight) : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>
  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<2x3x3x2xf32>
  %q_bias = "quantfork.qcast"(%bias) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>

    %cst_0 = "tf.Const"() {value = dense<[-2.000000e+00, 3.000000e+00]> : tensor<2xf32>} : () -> tensor<2xf32>
    %0 = "quantfork.qcast"(%cst) : (tensor<2x2x3x2xf32>) -> tensor<2x2x3x2x!quant.uniform<i8<-127:127>:f32:3, {4.000000e-03,5.000000e-03}>>

  %output, %min, %max, %histogram = "tf.CustomAggregator"(%arg0) <{calibration_method = 5 : i32, id = "0", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x3x4x3xf32>)...

  %w = arith.constant dense<-1.0> : tensor<32x3x3x3xf32>
  %3 = "tfl.quantize"(%w) {qtype = tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.1>>} : (tensor<32x3x3x3xf32>) -> tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.1>>
  %4 = "tfl.dequantize"(%3) : (tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.1>>) -> tensor<32x3x3x3xf32>