bool is_new_function = visited_functions.insert(function).second;
  if (!is_new_function) return absl::OkStatus();

  auto function_name = FindFunctionName(configs, function);

  // TODO(fengliuai): use a small flib_def to reduce overhead
  absl::flat_hash_set<Node*> control_ret_nodes;
  TF_ASSIGN_OR_RETURN(
      auto sub_graph,
      Exporter::Convert(configs, tf_dialect, symbol_table, function, flib_def,

	ALDGR
	ALGDR

	// convert from int32/int64 to float/float64
	ACEFBRA
	ACDFBRA
	ACEGBRA
	ACDGBRA

	// convert from float/float64 to int32/int64
	ACFEBRA
	ACFDBRA
	ACGEBRA
	ACGDBRA

	// convert from uint32/uint64 to float/float64
	ACELFBR
	ACDLFBR
	ACELGBR
	ACDLGBR

	// convert from float/float64 to uint32/uint64
	ACLFEBR
	ACLFDBR

    VLOG(2) << "MLIR module is not updated. Using the original graph. "
            << "Do not convert mlir module back to graph";
    return absl::OkStatus();
  }
  GraphExportConfig export_config;
  absl::flat_hash_set<Node*> control_ret_nodes;

  timings.Reset({kTfMlirCategory, "convert_mlir_to_graph"});
  // Some or all passes are enabled. Convert MLIR module and return back
  // resulted graph.

#include "tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h"
#include "xla/mlir_hlo/mhlo/IR/hlo_ops.h"

namespace mlir {
namespace odml {

// Convert updates into canonical form as expected by tf.scatter ops.
//
// tf.scatter expects `update_window_dims` to be the trailing dimensions.
//
// To support scatter ops generated by numpy-like slice updates:

	colName      []string // used by CREATE, INSERT, SELECT (selected columns)
	colType      []string // used by CREATE
	colValue     []any    // used by INSERT (mix of strings and "?" for bound params)
	placeholders int      // used by INSERT/SELECT: number of ? params

	whereCol []boundCol // used by SELECT (all placeholders)

	placeholderConverter []driver.ValueConverter // used by INSERT
}

			return nil
		}

		// attempt to convert the object to an external version first.
		target, ok := context.(GroupVersioner)
		if !ok {
			return fmt.Errorf("unable to convert the internal object type %T to Unstructured without providing a preferred version to convert to", in)
		}
		// Convert is implicitly unsafe, so we don't need to perform a safe conversion

  llvm::SmallSet<int, 8> read_array;
  for (const Attribute& attr : op.getDeviceVarReadsIndices()) {
    read_array.insert(mlir::cast<IntegerAttr>(attr).getInt());
  }
  llvm::SmallSet<int, 8> update_array;
  for (const Attribute& attr : op.getDeviceVarUpdatesIndices()) {
    update_array.insert(mlir::cast<IntegerAttr>(attr).getInt());
  }

  for (auto& arg : op->getOpOperands()) {
    Value v = arg.get();

	for _, pod := range pods {
		podFullName := kubecontainer.GetPodFullName(pod)
		// This logic relies on a static pod and its mirror to have the same name.
		// It is safe to type convert here due to the IsMirrorPod guard.
		if kubetypes.IsMirrorPod(pod) {
			mirrorPodUID := kubetypes.MirrorPodUID(pod.UID)
			pm.mirrorPodByUID[mirrorPodUID] = pod
			pm.mirrorPodByFullName[podFullName] = pod

            byte[] manifestBytes;
            if (DEFAULT_CONTENT_CHARSET.equals(contentCharset)) {
                manifestBytes = buffer.toByteArray();
            } else {
                // Convert the UTF-8 manifest bytes to the requested content charset
                manifestBytes = buffer.toString(DEFAULT_CONTENT_CHARSET).getBytes(contentCharset);
            }
            outputStream.write(manifestBytes);

            &build_ops_flags->tf_xla_print_cluster_outputs,
            "If true then insert Print nodes to print out values produced by "
            "XLA clusters."),
       Flag("tf_xla_check_cluster_input_numerics",
            &build_ops_flags->tf_xla_check_cluster_input_numerics,
            "If true then insert CheckNumerics nodes to check all cluster "
            "inputs."),