namespace tensorflow {
namespace image_format {

absl::Status ReadSavedModel(const std::string& file_prefix,
                            SavedModel* saved_model_proto) {
  LOG(INFO) << "Reading SavedModel from: " << file_prefix;

#if defined(PLATFORM_WINDOWS) || defined(__APPLE__)
  const std::string saved_model_pb_path = absl::StrCat(file_prefix, ".pb");
  TF_ASSIGN_OR_RETURN(
      bool saved_model_pb_exists,

#define IS_OSS false

namespace tensorflow {
namespace image_format {

// Reads the SavedModel proto from {file_prefix}{.pb|.cpb}.
// Returns a failure status when the SavedModel file does not exist.
absl::Status ReadSavedModel(const std::string& file_prefix,
                            SavedModel* saved_model_proto);

// Writes the SavedModel proto to a file or to string. If the proto is < the

      module = module_match.group()

      # Substitute all the function templates.
      out = re.split(file_prefix, src_file)
      if len(out) != 2:
        raise ValueError('The file name must start with {}'.format(file_prefix))
      tag = out[1][:-5]  # the last five values = ".mlir"
      module = _substitute_for_loop_template(module)

  // not expected to be persistent (usually a temporary directory). When
  // fetching the restore op (see `restore_node_name`), this value is provided
  // to the "file_prefix" tensor to identify the checkpoint directory.
  string checkpoint_dir = 5;

  // Function name -> function alias mapping. This associates the quantized

    absl::string_view file_prefix) {
  Env* env = Env::Default();
  std::vector<string> file_names;
  TF_RETURN_IF_ERROR(
      env->GetChildren(tensorflow::testing::TmpDir(), &file_names));

  bool altered = false;
  for (const auto& file_name : file_names) {
    if (absl::EndsWith(file_name, ".pb") &&
        absl::StartsWith(file_name, file_prefix)) {
      XlaSerializedCacheEntry entry;

  // found, returns NOT_FOUND error.
  Status AlterPersistentCacheEntryHloModuleNames(
      absl::string_view persistent_cache_dir_path,
      absl::string_view file_prefix = "xla_compile_cache");

 private:
  JitCompilationListener* listener_;
};

}  // namespace tensorflow

    tf_executor.graph {
      tf_executor.fetch
    }
    return
  }
// A new argument corresponding to the "file_prefix" should be created.
// Also checks that tf.entry_function is not created.
// CHECK: func.func @main(%[[ARG:.*]]: tensor<!tf_type.string> {tf_saved_model.index_path = ["file_prefix"]}) attributes {tf_saved_model.exported_names = ["main"]}
}

// -----

// Creates a pass that moves & merges the "@tf_quant__save" function to "@main"
// function. A new `IdentityOp` will be created. It will have control dependency
// to the save function and returns the file_prefix argument (typed
// `tensor<!tf_type.string>`). The file_prefix argument, which can be identified
// if the "tf_saved_model.index_path" attribute has "__tf_file_prefix", will be
// reused if it already exist in @main. Otherwise a new file prefix argument

// when the conversion fails.
//
// * `checkpoint_dir` is the directory where checkpoints where variable values
// are stored. This value will be fed to the "file_prefix" tensor to restore the
// variables.
// * `function_aliases` maps the actual function name to the function alias.
// This associates the quantized functions to the original functions' aliases.

}

// Dumps the MLIR module to disk.
// This require the TF_DUMP_GRAPH_PREFIX to be set to a path that exist (or can
// be created).
static void DumpModule(mlir::ModuleOp module, std::string file_prefix) {
  std::string prefix = GetDumpDirFromEnvVar();
  if (prefix.empty()) return;

  auto* env = tensorflow::Env::Default();
  auto status = env->RecursivelyCreateDir(prefix);
  if (!status.ok()) {