Implementation:Tensorflow Serving Session Bundle Util

Knowledge Sources	Tensorflow_Serving
Domains	Model Serving, Session Bundle, Migration
Last Updated	2026-02-13 00:00 GMT

Overview

Provides utility functions for migrating from the deprecated SessionBundle format to SavedModelBundle, including signature conversion, bundle conversion, and dual-format loading.

Description

The Session Bundle Util module (in the session_bundle namespace) provides backward-compatibility functions for working with the legacy SessionBundle format alongside the current SavedModel format. Key functions include:

Bundle Conversion:

• ConvertSignaturesToSignatureDefs: Converts legacy Signatures in a MetaGraphDef into the modern SignatureDef format.
• ConvertSessionBundleToSavedModelBundle: Converts a SessionBundle into a SavedModelBundle, enabling legacy models to be used with the SavedModel-based serving infrastructure.

Dual-Format Loading:

• LoadSessionBundleOrSavedModelBundle (two overloads): Attempts to load from either a SessionBundle path or a SavedModel path. Sets an optional is_session_bundle output flag indicating whether the model was up-converted from the legacy format. The second overload also supports loading SavedModel configuration from assets.extra/saved_model_config.pb.

Legacy Operations:

• LoadSessionBundleFromPathUsingRunOptions: Loads a SessionBundle with custom RunOptions for restore and init operations (marked as EXPERIMENTAL).
• SetSignatures: Sets Signatures in a MetaGraphDef.
• GetClassificationSignature / GetRegressionSignature: Extracts classification or regression signatures from the default signature in a MetaGraphDef.
• RunClassification / RunRegression: Executes classification or regression using legacy signature formats, validating input/output batch size consistency.
• GetNamedSignature: Retrieves a named Signature from a MetaGraphDef.

Usage

Use these utilities when migrating from SessionBundle to SavedModel format, or when serving infrastructure needs to support both formats simultaneously. The LoadSessionBundleOrSavedModelBundle function is particularly useful for backward-compatible model loading.

Code Reference

Source Location

• Repository: Tensorflow_Serving
• File: tensorflow_serving/session_bundle/session_bundle_util.h (lines 1-121)

Signature

namespace session_bundle {

Status ConvertSignaturesToSignatureDefs(MetaGraphDef* meta_graph_def);

Status ConvertSessionBundleToSavedModelBundle(
    SessionBundle& session_bundle, SavedModelBundle* saved_model_bundle);

Status LoadSessionBundleOrSavedModelBundle(
    const SessionOptions& session_options, const RunOptions& run_options,
    const string& export_dir, const std::unordered_set<string>& tags,
    SavedModelBundle* bundle, bool* is_session_bundle = nullptr);

Status RunClassification(const ClassificationSignature& signature,
                         const Tensor& input, Session* session,
                         Tensor* classes, Tensor* scores);

Status RunRegression(const RegressionSignature& signature,
                     const Tensor& input, Session* session,
                     Tensor* output);

}  // namespace session_bundle

Import

#include "tensorflow_serving/session_bundle/session_bundle_util.h"

I/O Contract

Inputs

Name	Type	Required	Description
session_options	SessionOptions	Yes	TensorFlow session configuration
run_options	RunOptions	Yes	Options for session run operations during loading
export_dir	string	Yes	Path to the model export directory (SessionBundle or SavedModel format)
tags	std::unordered_set<string>	Yes	Tags identifying which MetaGraphDef to load

Outputs

Name	Type	Description
bundle	SavedModelBundle*	Loaded (and possibly converted) SavedModelBundle
is_session_bundle	bool*	Set to true if the model was loaded from legacy SessionBundle format
return	Status	OK on success; error if loading or conversion fails

Usage Examples

Loading Either Format

SavedModelBundle bundle;
bool is_session_bundle;
Status status = session_bundle::LoadSessionBundleOrSavedModelBundle(
    SessionOptions(), RunOptions(), "/path/to/model",
    {"serve"}, &bundle, &is_session_bundle);
if (is_session_bundle) {
  LOG(INFO) << "Model was loaded from legacy SessionBundle format";
}

Related Pages

• Principle:Tensorflow_Serving_Session_Bundle_Migration