tf_executor.graph {
      %control = tf_executor.island {
        "tf.StatefulPartitionedCall"(%arg0) <{config = "", config_proto = "", executor_type = "", f = @partitioned}> : (tensor<*x!tf_type.resource>) -> ()
        tf_executor.yield
      }
      tf_executor.fetch %control : !tf_executor.control
    }

// -----

// This test makes sure functions with tf._noinline=true is not inlined.

module {
  func.func @stateful_partitioned_call(%arg0: tensor<1x2x2x3xf32>) -> (tensor<1x2x2x3xf32>) {
    %0 = "tf.StatefulPartitionedCall"(%arg0) <{
      config = "", config_proto = "", executor_type = "", f = @some_func
    }> {
      _collective_manager_ids = [], device = ""
    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>

def XlaClusterFormationPass : Pass<"tf-xla-cluster-formation", "ModuleOp"> {
  let summary = "Encapsulate partitioned calls within a Cluster op";
  let description = [{
    This pass clusters `tf.PartitionedCall` and `tf.StatefulPartitionedCall`
    with `_xla_compile_device_type` attribute into a `tf_device.cluster`.
    Notice this pass will only rewrite the outermost call if there are nested

  let summary = "Rewrites partition calls into Xla launch ops to make the attached function run on XLA.";

  let description = [{
    This pass rewrites `tf.PartitionedCall` and `tf.StatefulPartitionedCall`
    operations with `_xla_compile_device_type` attribute in a
    `tf_device.cluster` into `tf.XlaLaunch` operations. This makes the attached
    function execute with XLA. `tf.XlaLaunch` requires resource-type arguments

bool IsNonReplicatedGraph(const Graph& graph,
                          const FunctionLibraryDefinition* function_library) {
  auto predicate = [](const Graph& graph) {
    const std::string kStatefulPartitionedCallOp = "StatefulPartitionedCall";
    for (const Node* node : graph.nodes()) {
      auto node_op = node->type_string();
      if (node_op == kStatefulPartitionedCallOp) {
        // Functions called by StatefulfulPartitionedCall ops with

  pm.addNestedPass<mlir::func::FuncOp>(
      mlir::quant::stablehlo::CreateConvertTFQuantOpsToMHLOPass());
  pm.addPass(mlir::createCanonicalizerPass());

  // TF -> StableHLO legalization.
  // Skip StatefulPartitionedCall to preserve aliased functions.
  mlir::odml::AddLegalizeTFToStablehloPasses(pm, /*skip_quantization_ops=*/true,
                                             /*skip_resize=*/false,

  // CHECK: return %[[SUBGRAPH_1]] : tensor<1024x3xf32>
  // CHECK: }
}

// -----

// main function contains PartitionedCall and StatefulPartitionedCall ops which
// is used to preserve aliased functions. This test make sure stablehlo ops in
// each PartitionedCall functions are lifted.

  // we don't have any evidence that choosing a stateless partitioned call helps
  // for performance.
  ops::StatefulPartitionedCall call(
      root.WithOpName("stateful_partitioned_call"), args, n->output_types(),
      func, ops::StatefulPartitionedCall::Attrs{}.ConfigProto(config_string));

  for (const Edge* e : n->in_edges()) {
    if (e->IsControlEdge()) {

//         dtype: DT_FLOAT
//         tensor_shape: {
//         }
//       }
//     }
//     outputs: {
//       key  : "output_0"
//       value: {
//         name : "StatefulPartitionedCall:0"
//         dtype: DT_FLOAT
//         tensor_shape: {
//         }
//       }
//     }
//     method_name: "tensorflow/serving/predict"
//   }
// }

  }
}

TF::StatefulPartitionedCallOp MakeFuncCaller(
    mlir::OpBuilder& builder, const Location& loc, func::FuncOp func,
    const llvm::SetVector<Value>& operands) {
  // Constructs a tf.StatefulPartitionedCall to the function provided in 'func'
  // using the operands in 'operands'. Assumes the insertion point on builder is
  // already set.
  auto symbol =