%4 = "quantfork.stats"(%2) {layerStats = dense<[-1.36523, 3.57373]> : tensor<2xf32>} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    %5 = "chlo.broadcast_add"(%4, %3) : (tensor<1x2x2x2xf32>, tensor<2xf32>) -> tensor<1x2x2x2xf32>
    %6 = "quantfork.stats"(%5) {layerStats = dense<[-1.31055, 2.62842]> : tensor<2xf32>} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>

  %fill = "tf.Fill"(%cst_0, %cst) : (tensor<1xi64>, tensor<f32>) -> tensor<1xf32>
  tf_device.replicate([%0, %fill] as %arg_r0: tensor<1xf32>) {n = 2 : i32} {
    %1 = "tf_device.launch"() <{device = "TPU_REPLICATED_HOST_0"}> ({
      %2 = "tf.Identity"(%arg_r0) : (tensor<1xf32>) -> tensor<1xf32>
      tf_device.return %2 : tensor<1xf32>
    }) : () -> tensor<1xf32>

  %2 = "tfl.add"(%arg0, %arg3) {fused_activation_function = "RELU6", per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, per_device_costs = {CPU = 2.0 : f32, GPU = -1.0 : f32}, values_count = 2 : i32, tac.device = "CPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>

// CHECK-NEXT:   "tf.LayerNorm"(%arg4, %arg5, %arg6, %arg7) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
// CHECK-NEXT:   "tfl.yield"
// CHECK-NEXT: }) {_tfl_quant_trait = "fully_quantizable", device = ""} :

func.func @transpose_simple_1d() -> tensor<2xf32> {
  %0 = stablehlo.constant dense<[0.000000e+0, 1.000000e+0]> : tensor<2xf32>
  %1 = stablehlo.transpose %0, dims = [0] : (tensor<2xf32>) -> tensor<2xf32>
  return %1 : tensor<2xf32>
}
// CHECK-DAG: %[[CONST_0:.+]] = stablehlo.constant dense<[0.000000e+00, 1.000000e+00]> : tensor<2xf32>
// CHECK-NOT: transpose
// CHECK: return %[[CONST_0]] : tensor<2xf32>

// -----

  %1 = stablehlo.constant dense<[4.0, 6.0]> : tensor<2xf32>
  %2 = stablehlo.divide %0, %1 : tensor<2xf32>
  return %2 : tensor<2xf32>
}

// CHECK-LABEL: foldDivLHSSplat
// CHECK: stablehlo.constant dense<[5.000000e-01, 0.333333343]> : tensor<2xf32>

// -----

func.func @foldDivRHSSplat() -> tensor<2xf32> {
  %0 = stablehlo.constant dense<[4.0, 6.0]> : tensor<2xf32>
  %1 = stablehlo.constant dense<2.0> : tensor<2xf32>

  %2 = "stablehlo.concatenate"(%0, %1) {dimension = 0 : i64} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
  %3 = "quantfork.stats"(%2) {layerStats = dense<[-0.83811146, 2.4960899]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>
  %4 = "quantfork.stats"(%arg2) {layerStats = dense<[-1.5726943, 1.07351148]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>

^bb0(%arg0: tensor<40x37xf32>, %arg1: tensor<40x37xf32>):
  %cst = arith.constant dense<1.0> : tensor<40xf32>
  %0:2 = "tfl.fully_connected"(%arg0, %arg1, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<40x37xf32>, tensor<40x37xf32>, tensor<40xf32>) -> (tensor<40x40xf32>, tensor<40x40xf32>)
  func.return %0 : tensor<40x40xf32>

func.func @main(%arg0: tensor<32x4x4x128xf32>, %arg1: tensor<1x32x42x128xf32>, %arg2: tensor<4xi32>) -> tensor<1x64x84x32xf32> {
  %0 = "tfl.custom"(%arg0, %arg1, %arg2) {custom_code = "Convolution2DTransposeBias", custom_option = #tfl<const_bytes : "0x010000000200000002000000">} : (tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<4xi32>) -> tensor<1x64x84x32xf32>
  func.return %0 : tensor<1x64x84x32xf32>
}
// CHECK-LABEL: main

func.func @main(%arg0: tensor<32x4x4x128xf32>, %arg1: tensor<1x32x42x128xf32>, %arg2: tensor<4xi32>) -> tensor<1x64x84x32xf32> {
  %0 = "tfl.custom"(%arg0, %arg1, %arg2) {custom_code = "Convolution2DTransposeBias", custom_option = #tfl<const_bytes : "0x010000000200000002000000">} : (tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<4xi32>) -> tensor<1x64x84x32xf32>
  func.return %0 : tensor<1x64x84x32xf32>
}
// CHECK-LABEL: main