FunctionDef fdef = XTimesY();
  (*fdef.mutable_attr())["_XlaMustCompile"] = BoolAttr(true);
  Init({fdef});
  XlaKernelCreator xla_kernel_creator;
  auto callsite =
      ToNodeProperties(R"pb(
        name: 'XTimesY' op: 'XTimesY' input: 'a' input: 'b'
      )pb");
  (*(callsite->node_def.mutable_attr()))["_XlaMustCompile"] = BoolAttr(true);

  // Note: need to set attribute on the created node.

// attributes into unified attributes. For example, A _tpu_replicate=X
// should be replaced with _xla_compile_device_type=TPU and
// _replication_info=X attributes by the conversion. An _XlaMustCompile=true
// should be replaced with _xla_compile_device_type with the value of device
// attribute.

#include "llvm/ADT/StringRef.h"
#include "llvm/Support/Debug.h"
#include "mlir/IR/Builders.h"  // from @llvm-project

#include "tensorflow/core/framework/function.h"

namespace tensorflow {

// Checks if a graph or reachable functions in the library have any
// StatefulPartitionedOps with _XlaMustCompile=true. The function library will
// be skipped if nullptr is provided.
bool IsSupportedByNonReplicatedBridge(
    const Graph& graph, const FunctionLibraryDefinition* function_library);

  %0 = "tf.StatefulPartitionedCall"(%arg0) {_XlaMustCompile = true, config = "", config_proto = "", device = "/device:CPU:0", executor_type = "", f = @stateful_pcall_func} : (tensor<i32>) -> (tensor<i32>)
  func.return %0 : tensor<i32>
}

  %1 = "tf.AddV2"(%arg0, %arg0) {} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  func.return %1 : tensor<i32>
}

func.func private @xla_func_0(%arg0: tensor<1x3xf32>, %arg1: tensor<1x3xf32>) -> tensor<1x3xf32> attributes {tf._XlaMustCompile = true, tf._noinline = true, tf._original_func_name = "should_not_be_used"} {
  %1 = "tf.AddV2"(%arg0, %arg1) : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>

using mlir::func::FuncOp;

#define GEN_PASS_DEF_XLACLUSTERFORMATIONPASS
#include "tensorflow/compiler/mlir/tf2xla/internal/passes/clustering_passes.h.inc"

// Outlines partitioned call ops with `_XlaMustCompile` to device clusters.
struct XlaClusterFormationPass
    : public impl::XlaClusterFormationPassBase<XlaClusterFormationPass> {
  void runOnOperation() override;
};

void EncapsulatePartitionedCall(Operation *call_op,

    for (const Node* node : graph.nodes()) {
      auto node_op = node->type_string();
      if (node_op == kStatefulPartitionedCallOp) {
        // Functions called by StatefulfulPartitionedCall ops with
        // _XlaMustCompile=true are compiled by XLA.
        auto attr = node->attrs().FindByString(std::string(kMustCompileAttr));
        if (attr != nullptr && attr->b() == true) {
          return true;
        }
      }
    }

    if (isa<TPUPartitionedCallOp>(call)) return false;
    if (isa<RemoteCallOp>(call)) return false;
    // Maintain inlining for  `tf.function`s with jit_compile option.
    if (callable->hasAttr("tf._XlaMustCompile")) return true;
    auto noinline_attr_name = absl::StrCat("tf.", tensorflow::kNoInlineAttr);
    if (auto noinline_attr =
            callable->getAttrOfType<BoolAttr>(noinline_attr_name))

  return %x : tensor<i32>
}

// -----

// Test for XlaLaunch

func.func private @xla_func_0(%arg0: tensor<1x3xf32>, %arg1: tensor<1x3xf32>) -> tensor<1x3xf32> attributes {tf._XlaMustCompile = true, tf._noinline = true, tf._original_func_name = "should_not_be_used"} {
  %1 = "tf.AddV2"(%arg0, %arg1) {__op_key = 0: i32} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %1 : tensor<1x3xf32>
}

      // tf.StatefulPartitionedCall ops with and without
      // _tpu_replicate attributes may exist in the same graph. Ops without
      // the attribute but with _XlaMustCompile=true would have
      // _xla_compile_device_type="" after
      // CanonicalizeCompileAndReplicateAttributesPass. Skip empty value here.
      if (!device_type_attr.getValue().empty()) {