%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 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} {

  %4 = stablehlo.broadcast_in_dim %1, dims = [3] : (tensor<4xf32>) -> tensor<1x3x3x4xf32>
  %5 = stablehlo.add %3, %4 : tensor<1x3x3x4xf32>
  %6 = stablehlo.maximum %5, %2 : tensor<1x3x3x4xf32>
  func.return %6: tensor<1x3x3x4xf32>
}
// CHECK: %[[CONST_0:.*]] = stablehlo.constant dense<2.000000e+00>

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>

    %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>

  %2 = "tfl.select"(%cst_false, %arg0, %arg1) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>
  %3 = "tfl.select_v2"(%cst_false, %arg0, %arg1) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>
  func.return %0, %1, %2, %3 : tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>

  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>

    %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}