flutter_litert_flex

FlexDelegate (SELECT_TF_OPS) addon for flutter_litert. Automatically downloads and bundles the TensorFlow Lite Flex delegate native library at build time on iOS, macOS, Linux and Windows.

Android is not supported. flutter_litert 4.0.0 runs on LiteRT 2.x on Android, and LiteRT 2.x has no flex delegate, so flutter_litert_flex 2.0.0 has no Android side. Android apps that need flex can stay on flutter_litert: ^3.10.0 with flutter_litert_flex: 1.3.0, or reconvert the model to LiteRT built-in ops.

When do you need this?

The Flex delegate is required when your .tflite model uses TensorFlow ops that aren't available as TFLite builtins. Common cases:

• On-device training with convolutional or batch-normalized layers (gradient ops like Conv2DBackpropFilter)
• Checkpoint-based persistence using tf.raw_ops.Save/Restore
• Models converted with SELECT_TF_OPS enabled

If your model only uses TFLITE_BUILTINS, you don't need this package.

Installation

dependencies:
  flutter_litert: ^4.0.0
  flutter_litert_flex: ^2.0.0

That's it. The native libraries are downloaded automatically on the first build.

Compatibility

Use the matching flutter_litert version:

flutter_litert	flutter_litert_flex	Flex platforms
4.0.0 or newer	2.0.0	iOS 15+, macOS, Linux, Windows
3.10.x	1.3.0	Android, iOS, macOS, Linux, Windows

flutter_litert 4.0.0 and flutter_litert_flex 2.0.0 need iOS 15.0 or newer.

iOS + Swift Package Manager: flutter_litert_flex must be used alongside flutter_litert; the FlexDelegate links against core TFLite symbols that flutter_litert provides. Like flutter_litert, its xcframework ships arm64 device + arm64 simulator slices, so iOS simulator builds require an Apple Silicon Mac; the x86_64 (Intel) simulator is not supported under SPM. CocoaPods builds are unaffected.

Usage

import 'package:flutter_litert/flutter_litert.dart';

final options = InterpreterOptions();
options.addDelegate(await FlexDelegate.create());
final interpreter = await Interpreter.fromAsset('model.tflite', options: options);

FlexDelegate.isAvailable tells you whether the library was bundled. It is always false on Android, where FlexDelegate.create() throws an UnsupportedError.

Platform details

Platform	Library	Size	Mechanism
iOS	TensorFlowLiteFlex.xcframework	~492 MB	CocoaPods or SPM (static framework)
macOS	libtensorflowlite_flex-mac.dylib	~123 MB	CocoaPods or SPM (bundled resource)
Linux	libtensorflowlite_flex-linux.so	~333 MB	CMake download
Windows	libtensorflowlite_flex-win.dll	~227 MB	CMake download

Libraries are downloaded once and cached in the plugin directory. Subsequent builds use the cached copy.

Credits

Based on TensorFlow Lite v2.20.0. See flutter_litert for the base inference library.