%3 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1
  %4 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1

    %0 = "tf.Const"() {device = "/device:CPU:0", value = dense<0> : tensor<i64>} : () -> tensor<i64>
    %1 = "tf.Const"() {device = "/device:CPU:0", value = dense<5> : tensor<i64>} : () -> tensor<i64>
    %2 = "tf.Const"() {device = "/device:CPU:0", value = dense<1> : tensor<i64>} : () -> tensor<i64>

        // We assign default device to ops with device attribute that is empty.
        if (device_attr.getValue().empty()) {
          op->setAttr(kDeviceAttr, builder.getStringAttr(default_device_));
        }
      } else if (op->getDialect() == tf) {
        // Assign default device to all ops in Tensorflow dialect that do not
        // have device attribute.

  // CHECK: tac.cost = 1.000000e+03
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  // CHECK: tac.cost = 1.000000e+03
  %1 = "tfl.mul"(%0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  func.return %1 : tensor<10x10x10xf32>
}

  if (!device_attr || device_attr.getValue().empty()) return true;

  DeviceNameUtils::ParsedName device;
  if (!DeviceNameUtils::ParseFullName(device_attr.getValue().str(), &device))
    return false;

  // We can't use GPU if operation explicitly placed on non-GPU device.
  return !device.has_type || device.type == ::tensorflow::DEVICE_GPU;
}

}  // namespace TF

  %cst = "tf.Const"() {device = "", value = dense_resource<__elided__> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %cst_0 = "tf.Const"() {device = "", value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %cst_1 = "tf.Const"() {_tpu_replicate = "cluster", device = "", value = dense<[0, 2, 3, 1]> : tensor<4xi32>} : () -> tensor<4xi32>

      %outputs, %control = "tf_executor.island"() ({
        %0 = "tf.Const"() {device = "", value = dense<0> : tensor<i32>} : () -> tensor<i32>
        "tf_executor.yield"(%0) : (tensor<i32>) -> ()
      }) : () -> (tensor<i32>, !tf_executor.control)

  // CHECK: tfrt_fallback_async.set_resource {{%.*}}, {{%.*}} {device = "cpu", index = 1 : i64}
  // CHECK: tfrt_fallback_async.get_resource {{%.*}} {_tfrt_cost = 1 : i64, device = "cpu", indices = [0, 1]}
  %ch1 = tfrt_fallback_async.set_resource %ch0, %ra {device = "cpu", index = 0 : i64}
  %ch2 = tfrt_fallback_async.set_resource %ch1, %rb {device = "cpu", index = 1 : i64}

    %3:3 = tf_executor.Merge %2#0, %0#0 : tensor<*x!tf_type.int32ref> {device = "", N = 2, T = "tfdtype$DT_INT32"} loc("while/Merge")
    %4:2 = tf_executor.island(%3#2) wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_INT32", value = dense<10> : tensor<i32>} : () -> tensor<i32> loc("while/Less/y")

  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "GPU"} : (tensor<10x10x10xf32>, tensor<f32>) -> tensor<10x10x10xf32>
  func.return %0 : tensor<10x10x10xf32>
}

// -----

// CHECK: tac.cost = 2.000000e+03
func.func @func_0_CPU(%arg0: tensor<10x10x10xf32>, %arg1: tensor<10xf32>) -> tensor<10x10x10xf32> attributes {tac.device = "CPU", tac.interface_name = "func_0"} {