%2 = func.call @fold_all_test(%arg0, %0, %1) : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>
  func.return %2 : tensor<256x30x30x16xf32>
}

// ALL-LABEL: @fold_all_test
func.func @fold_all_test(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>) -> tensor<256x30x30x16xf32> {

  pass_config.legalize_custom_tensor_list_ops =
      toco_flags.legalize_custom_tensor_list_ops();
  // Disable the unfolding of the 16x16 TF::BatchMatMulOp to avoid the
  // conversion to an unsupported 16x16 TFL::FullyConnectedOp.
  if (toco_flags.inference_type() == toco::IODataType::QUANTIZED_INT16) {
    pass_config.unfold_batch_matmul = false;
  }

// CHECK:  return
}

func.func @mul_i16(%arg0: tensor<14xi16>, %arg1: tensor<14xi16>) -> tensor<14xi16> {
  %0 = "tf.Mul"(%arg0, %arg1) : (tensor<14xi16>, tensor<14xi16>) -> tensor<14xi16>
  func.return %0: tensor<14xi16>

// CHECK-LABEL: mul_i16
// CHECK:  tfl.mul %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<14xi16>
// CHECK:  return
}

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<255.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

func.func @main() -> (tensor<1x2xf16>, tensor<2xf16>) {
  %graph:2 = tf_executor.graph {
    %0:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_HALF", value = dense<1.0> : tensor<1x2xf16>} : () -> tensor<1x2xf16> loc("const1")
    %1:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_HALF", value = dense<[1.0, 2.0]> : tensor<2xf16>} : () -> tensor<2xf16> loc("const2")

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<15.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

      Targs = [tensor<32xf32>, tensor<64xf32>, tensor<16xf32>],
      Tresults = [tensor<32xf32>, tensor<16xf32>]}
      : (tensor<32xf32>, tensor<64xf32>, tensor<16xf32>, tensor<2x!tf_type.string>) -> (tensor<32xf32>, tensor<16xf32>)
    tf_device.return %0#0, %0#1 : tensor<32xf32>, tensor<16xf32>
  }) {device = "/job:localhost/replica:0/task:0/device:TPU:0"} : () -> (tensor<32xf32>, tensor<16xf32>)
  // CHECK-NEXT: tf_device.return
  // CHECK-NEXT: })

  %0 = "tf.Const"() {value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : () -> tensor<4xui64>
  // CHECK: tfrt_fallback_async.const_dense_tensor  dense<1.000000e+00> : tensor<1xbf16>
  %1 = "tf.Const"() {device = "/device:CPU:0", value = dense<[1.0]> : tensor<1xbf16>} : () -> tensor<4xbf16>
  // CHECK: corert.executeop({{.*}}) "tf.Const"() {dtype = ui64, value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : 1

  %1 = "tfl.pseudo_const"() { value = dense<1.0> : tensor<16xf32>} : () -> tensor<16xf32>
  %2 = "tfl.conv_3d"(%arg0, %0, %1) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, dilation_d_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32, stride_d = 1 : i32} : (tensor<1x32x32x32x8xf32>, tensor<1x1x1x8x16x!quant.uniform<i8:f32, 1.0>>, tensor<16xf32>) -> tensor<1x32x32x32x16xf32>

  %0, %1, %2, %3, %4, %5 = "tf.FusedBatchNormV3"(%arg0, %arg1, %arg2, %arg3, %arg4) {data_format = "NHWC", device = "", epsilon = 1.000000e-03 : f32, exponential_avg_factor = 1.000000e+00 : f32, is_training = false} : (tensor<?x112x112x16xf16>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>) -> (tensor<?x112x112x16xf16>,...