func.func @variant_input_output_already_marked() {
  // CHECK: "tf.opA"
  // CHECK-SAME: _xla_outside_compilation = "0"
  // CHECK: "tf.opB"
  // CHECK-SAME: _xla_outside_compilation = "0"
  "tf_device.cluster"() ({
    %1= "tf.opA"() {_xla_outside_compilation = "0"} : () -> tensor<!tf_type.variant<tensor<f32>>>
    "tf.opB"(%1) {_xla_outside_compilation = "0"} : (tensor<!tf_type.variant<tensor<f32>>>) -> ()
    tf_device.return

    // CHECK-NEXT:   %[[A_OUT:.*]] = "tf.A"
    // CHECK-NOT:    _xla_outside_compilation
    // CHECK-NEXT:   %[[B_OUT:.*]] = "tf.B"(%[[A_OUT]])
    // CHECK-NOT:    _xla_outside_compilation
    // CHECK-NEXT:   %[[C_OUT:.*]] = "tf.C"(%[[B_OUT]], %arg0)
    // CHECK-NOT:    _xla_outside_compilation
    // CHECK-NEXT:   tf_device.return %[[C_OUT]]
    //
    // CHECK:      "tf_device.cluster"

    For example, the following tf_device.cluster with an op marked for `xla_outside_compilation`:

    ```mlir
    func @outside_compilation() -> tensor<f32> {
      %0 = "tf_device.cluster"() ( {
        %1 = "tf.Const"() {_xla_outside_compilation = "0", value = dense<1.0> : tensor<f32>} : () -> (tensor<f32>)
        %2 = "tf.Identity"(%1) {_xla_outside_compilation = "0"} : (tensor<f32>) -> (tensor<f32>)

tensor<1120x?xi32>, tensor<2xi32>) {\0A    %0 = \22tf.Reshape\22(%arg0, %arg1) {_xla_outside_compilation = \220\22} : (tensor<3360x?xi32>, tensor<3xi32>) -> tensor<3x1120x?xi32> loc(#loc9)\0A    %1:3 = \22tf.Split\22(%arg2, %0) {_xla_outside_compilation = \220\22} : (tensor<i32>, tensor<3x1120x?xi32>) -> (tensor<1x1120x?xi32>, tensor<1x1120x?xi32>, tensor<1x1120x?xi32>) loc(#loc10)\0A    %2 = \22tf.Reshape\22(%1#0, %arg3) {_xla_outside_compilation = \220\22} : (tensor<1x1120x?xi32>, tensor<2xi32>) -> tensor<1120x?xi32>...

  let summary = "Verifies that after Outside Compilation passes complete, there are no more _xla_outside_compilation attributes and no tf_device.launch ops.";

  let description = [{
    This pass takes iterates over ops and ensures that _xla_outside_compilation
    and tf_device.launch ops no longer exist and have been removed. This happens
    after Outside Compilation passes have completed.
  }];

  // Automatically marked ops for outside compilation have
  // `_xla_outside_compilation` attribute value of "auto" plus
  // an increasing counter.  Manually marked ops for outside compilation only
  // have an increasing counteri for the attribute value.  Therefore there is no
  // collision in
  // `_xla_outside_compilation` attribute between automatically and manually
  // marking ops.

#include "tensorflow/compiler/mlir/tensorflow/utils/tpu_rewrite_device_util.h"

namespace tensorflow {
namespace tf2xla {
namespace internal {

// This pass extracts a CPU computation cluster with `_xla_outside_compilation`
// annotation from the head or tail of a TPU cluster.

namespace {

using mlir::Block;
using mlir::BlockArgument;
using mlir::BoolAttr;
using mlir::ModuleOp;
using mlir::OpBuilder;

  auto opts = builder.opts();
  Node* const0 = ops::SourceOp("InputFloatOp", opts);
  Node* uncompilable_op = ops::UnaryOp("MissingKernel", const0, opts);
  uncompilable_op->AddAttr("_xla_outside_compilation", "0");
  GraphDef graph_def;
  TF_EXPECT_OK(builder.ToGraphDef(&graph_def));

  auto* flib_runtime = GetFunctionLibraryRuntime();

  // Outside compiled ops are considered by default..

// attribute on the `tf_device.launch`.
std::unique_ptr<OperationPass<func::FuncOp>> CreateLaunchToDeviceAttributePass(
    bool legacy_graph_export = true);

// Creates a pass to ensure that the `_xla_outside_compilation` and
// tf_device.launch op no longer exist after Outside Compilation is complete.
std::unique_ptr<OperationPass<func::FuncOp>>
CreateVerifyNoOutsideCompilationMarkersPass();