# CHECK-SAME: shared_name = "batch"
# CH#CK: sym_name = "predict0"

# CHECK: f = @inference_func[[post_fix]],
# CHECK-SAME: shared_name = "batch"
# CH#CK: sym_name = "predict1"

meta_graphs: {
  meta_info_def: {
    tags: ["serve", "tpu"]
  }
  graph_def: {
    node: {
      name: "input0"
      op: "Placeholder"
      attr: {
        key: "dtype"
        value: {

#include <utility>
#include <vector>

#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/protobuf/meta_graph.pb.h"

namespace tensorflow {
namespace saved_model {

std::string GetWriteVersion(const SavedModel& saved_model) {
  if (saved_model.meta_graphs_size() == 1 &&
      saved_model.meta_graphs()[0].has_object_graph_def()) {
    return "2";
  }
  return "1";
}

std::set<std::string> GetMapKeys(

        field_checksum = FingerprintCat64(field_checksum, hash);
      }
    }
  }
  return field_checksum;
}

inline RepeatedPtrField<FieldIndex> GraphDefFieldTags() {
  // SavedModel.meta_graphs[0].graph_def
  FieldIndex meta_graph_field_tag;
  meta_graph_field_tag.set_field(2);
  FieldIndex meta_graph_index_field_tag;
  meta_graph_index_field_tag.set_index(0);
  FieldIndex graph_def_field_tag;

  }

  // Build a map of node names to their corresponding nodes.
  //
  // See `GetGraphDefNode` for more details.
  const auto& nodes =
      tf_package.saved_model_proto_.meta_graphs(0).graph_def().node();
  for (const auto& node : nodes) {
    tf_package.graph_def_nodes_by_name_[node.name()] = &node;
  }
  return tf_package;
}

  saved_model_proto = saved_model_loader.parse_saved_model(saved_model_path)
  for meta_graph in saved_model_proto.meta_graphs:
    if any(
        node.op.startswith('FakeQuant') for node in meta_graph.graph_def.node
    ):
      return True
    for function in meta_graph.graph_def.library.function:
      if any(node.op.startswith('FakeQuant') for node in function.node_def):
        return True

};

// Restore variable and resources in the SavedModel export dir for the
// indicated metagraph.
// The recommended way to load a saved model is to call LoadSavedModel,
// which provides an already initialized Metagraph, Session, and DebugInfo.
Status RestoreSession(const RunOptions& run_options,
                      const MetaGraphDef& meta_graph, const string& export_dir,
                      std::unique_ptr<Session>* session);

  // Create a copy of `metagraph` which will be used and mutated for fingerprint
  // computation.
  FingerprintDef fingerprint_def;
  MetaGraphDef* metagraph = saved_model.mutable_meta_graphs(0);
  // Set fingerprint field #1.
  fingerprint_def.set_saved_model_checksum(HashSavedModel(saved_model));
  // Set fingerprint field #2.
  graph_regularization::SimpleDelete(*metagraph->mutable_graph_def());

Status ValidateSingleConcreteFunction(const SavedFunction& saved_function);

// Walks through the SavedObjectGraph in metagraph, and restores all nodes
// (except "UserDefinedObjects") with their corresponding type in
// "PartiallyRevivedObjects".
Status PartiallyReviveSavedModelObjects(const MetaGraphDef& metagraph,
                                        ImmediateExecutionContext* context,

                                const MetaGraphDef& meta_graph,
                                std::unique_ptr<Session>* session) {
  Session* session_p = nullptr;
  TF_RETURN_IF_ERROR(NewSession(session_options, &session_p));
  session->reset(session_p);
  TF_RETURN_IF_ERROR(ValidateSavedTensors(meta_graph.graph_def()));
  return (*session)->Create(meta_graph.graph_def());
}

                   const std::unordered_set<std::string>& tags) {
  tensorflow::MetaGraphDef meta_graph;
  TF_RETURN_IF_ERROR(tensorflow::ReadMetaGraphDefFromSavedModel(
      saved_model_path, tags, &meta_graph));

  absl::flat_hash_map<FunctionName, FunctionAlias> function_aliases(
      meta_graph.meta_info_def().function_aliases().begin(),
      meta_graph.meta_info_def().function_aliases().end());
  return function_aliases;
}