return %1 : tensor<1x3x4x2xf32>
  }

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

    return %2 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.convolution(%arg0, %arg1) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[0, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>

  func.return %dequant : tensor<1x3x4x2xf32>
}

// -----

// CHECK-LABEL: @dont_merge_quantization_followed_by_quantization
// CHECK-SAME: %[[ARG_0:.*]]: tensor<1x3x4x3xf32>
func.func @dont_merge_quantization_followed_by_quantization(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32> {
  // CHECK: %[[QUANT_ARG_0:.*]] = stablehlo.uniform_quantize %[[ARG_0]]

    %2 = "tf.Relu"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
    %3 = "tf.FakeQuantWithMinMaxArgs"(%2) {device = "", max = 4.000000e-01 : f32, min = -3.000000e-01 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

// CHECK: return %[[identity]] : tensor<1x3x2x2xf32>

func.func @cast_bf16_avg_pool_to_fp32(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.000000e+00> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>

// Remove TPU related ops.
func.func @tpu_conv(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
  %0 = "tf.TPUOrdinalSelector"() {device = ""} : () -> tensor<?xi32>
  %1 = "tf.TPUPartitionedCall"(%arg0, %0) {autotuner_thresh = 0 : i64, device = "", f = @tpu_func_0_optim0} : (tensor<1x3x4x3xf32>, tensor<?xi32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %2 : tensor<1x3x2x2xf32>
}

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %3 : tensor<1x3x2x2xf32>
// CHECK: tf.Conv2D
// CHECK-SAME: loc("Model/conv2d")
}

func.func @conv2d_with_valid_loc(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {

        : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32> loc(fused["Conv2D:", "test_opt_out"])
    %2 = "tf.BiasAdd"(%1, %cst_0) : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32> loc(fused["BiasAdd:", "model/bias_add"])
    %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
    return %3 : tensor<1x3x2x2xf32>
  }
}

    %0 = "quantfork.stats"(%arg1) {layerStats = dense<[-3.54062747, 0.54742622]> : tensor<2xf32>} : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2xf32>
    %1 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 2.824783]> : tensor<2xf32>} : (tensor<1x3x2x3xf32>) -> tensor<1x3x2x3xf32>

  %3 = "tf.BiasAdd"(%2, %cst_0) {data_format = "NHWC", device = ""} : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32>
  %4 = "tf.Relu"(%3) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  %5 = "quantfork.qcast"(%4) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0027450981093387976:-19>>
  %6 = "quantfork.dcast"(%5) : (tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0027450981093387976:-19>>) -> tensor<1x3x2x2xf32>