Flutter ONNX Runtime GenAI

A Flutter FFI plugin that wraps the Microsoft ONNX Runtime GenAI C-API for on-device multimodal inference.

✅ Tested Models

Model	Device	Status	Notes
Gemma 3 4B Instruct	Pixel 8a	✅ Working	INT4 quantized

Tested other models or devices? Please open a PR to add them!

📊 See detailed performance benchmarks →

🤝 Contributing & Help Wanted

We are looking for help to develop and test this package! Whether you have experience with Dart FFI, ONNX Runtime, or native mobile development (Swift/Kotlin/C++), your contributions are highly welcome.

If you'd like to help, please:

• Open an issue to discuss proposed changes.
• Submit pull requests for bug fixes or new features.
• Help with testing on different devices and documenting the results.

Features

• Vision-Language Models: Run models like Phi-3.5 Vision on-device
• Cross-Platform: Android and iOS support
• FFI Performance: Direct native library access via Dart FFI
• Async Inference: Background isolate support for non-blocking UI
• Streaming: Token-by-token output streaming

Requirements

• Flutter 3.3.0+
• Dart SDK 3.10.4+
• Android: API 24+ (NDK for building), arm64 device only (x86_64 emulator not included in pub.dev package)
• iOS: 13.0+ (Xcode for building), device only (simulator not included in pub.dev package)

Note: The pub.dev package only includes device libraries (Android arm64, iOS arm64) to stay under size limits. For simulator/emulator testing, build from source using the instructions below.

Installation

1. Add Dependency

dependencies:
  flutter_onnxruntime_genai: ^0.1.6

Or from Git for the latest development version:

dependencies:
  flutter_onnxruntime_genai:
    git:
      url: https://github.com/colymba/flutter_onnxruntime_genai.git

2. Set Up ONNX Runtime GenAI Submodule (Optional - for building from source)

cd flutter_onnxruntime_genai
git submodule add https://github.com/microsoft/onnxruntime-genai.git native_src/onnxruntime-genai
git submodule update --init --recursive

3. Build Native Libraries (Optional)

Prebuilt binaries for Android (arm64, x86_64) and iOS (Device & Simulator) are included in the package. You only need to build from source if you want to customize the build or use a different version:

./scripts/build_onnx_libs.sh all

Usage

Basic Inference

import 'package:flutter_onnxruntime_genai/flutter_onnxruntime_genai.dart';

// Create instance
final onnx = OnnxGenAI();

// Check library version
print('Library version: ${onnx.libraryVersion}');

// Verify model health
final status = await onnx.checkNativeHealthAsync('/path/to/model');
if (status == HealthStatus.success) {
  print('Model is ready!');
} else {
  print('Error: ${HealthStatus.getMessage(status)}');
}

// Run multimodal inference (text + image)
try {
  final result = await onnx.runInferenceAsync(
    modelPath: '/path/to/phi-3.5-vision',
    prompt: 'Describe this image in detail.',
    imagePath: '/path/to/image.jpg',
  );
  print('Generated: $result');
} on OnnxGenAIException catch (e) {
  print('Inference error: $e');
}

Text-Only Inference

final result = await onnx.runTextInferenceAsync(
  modelPath: '/path/to/model',
  prompt: 'What is the capital of France?',
  maxLength: 256,
);

Streaming Output

final streamer = OnnxGenAIStreamer();

await for (final token in streamer.streamInference(
  modelPath: '/path/to/model',
  prompt: 'Tell me a story about a brave knight.',
)) {
  stdout.write(token); // Print tokens as they're generated
}

⚡ Performance Optimization with Execution Providers

For better performance on mobile devices, you can configure execution providers at runtime. This allows you to:

• Use hardware-accelerated inference (XNNPACK, QNN, CoreML)
• Override settings without modifying genai_config.json

// Use XNNPACK for optimized ARM inference on mobile
final result = await onnx.runInferenceWithConfigAsync(
  modelPath: '/path/to/model',
  prompt: 'Hello, how are you?',
  providers: ['XNNPACK'],  // Use XNNPACK execution provider
);

Supported Execution Providers

Provider	Platform	Description	Best For
XNNPACK	Android/iOS	Optimized ARM NEON kernels	Mobile ARM devices
QNN	Android	Qualcomm Hexagon NPU	Snapdragon devices
CoreML	iOS/macOS	Apple Neural Engine	Apple devices with ANE
SNPE	Android	Snapdragon Neural Processing Engine	Qualcomm devices
OpenVINO	Desktop	Intel optimized inference	Intel CPUs/GPUs
DML	Windows	DirectML GPU acceleration	Windows with DirectX 12

Note: cpu is NOT a valid provider name. CPU execution is the default fallback when no other provider is available. Thread configuration is done via the model's genai_config.json file.

Architecture

flutter_onnxruntime_genai/
├── native_src/
│   └── onnxruntime-genai/      # Git submodule
├── src/
│   ├── include/
│   │   └── ort_genai_c.h       # C-API header (stub)
│   ├── flutter_onnxruntime_genai.cpp
│   ├── flutter_onnxruntime_genai.h
│   └── CMakeLists.txt
├── android/
│   ├── build.gradle
│   └── src/main/jniLibs/       # Prebuilt .so files
├── ios/
│   ├── flutter_onnxruntime_genai.podspec
│   └── Frameworks/             # XCFramework
├── lib/
│   ├── flutter_onnxruntime_genai.dart
│   └── src/
│       └── ffi_bindings.dart   # Dart FFI layer
└── scripts/
    └── build_onnx_libs.sh      # Cross-compilation script

API Reference

OnnxGenAI

Main class for ONNX Runtime GenAI operations.

Method	Description
libraryVersion	Get the native library version
checkNativeHealth(path)	Verify model can be loaded
runInference(...)	Multimodal inference (blocking)
runTextInference(...)	Text-only inference (blocking)
runInferenceAsync(...)	Multimodal inference (background isolate)
runTextInferenceAsync(...)	Text-only inference (background isolate)
runInferenceWithConfigAsync(...)	Inference with custom execution providers (recommended)
runInferenceMultiWithConfigAsync(...)	Multi-image inference with custom providers (recommended)
createConfig(path)	Create a configuration object
destroyConfig(handle)	Free configuration resources
configClearProviders(handle)	Clear all providers from config
configAppendProvider(handle, name)	Add an execution provider
configSetProviderOption(...)	Set provider-specific options
getLastError()	Get last error message from native layer
shutdown()	Release native resources

HealthStatus

Status codes from checkNativeHealth:

Code	Constant	Meaning
1	success	Model loaded successfully
-1	invalidPath	NULL or empty path
-2	modelCreationFailed	Model creation failed
-3	tokenizerCreationFailed	Tokenizer creation failed

Important Notes

Threading

All synchronous inference methods (runInference, runTextInference) are blocking operations that can take seconds to minutes depending on the model and prompt.

DO NOT call these from the main UI isolate. Use the async variants which automatically run in a background isolate.

Android 15 Compatibility

The build script and CMake configuration include the critical 16KB page alignment flag (-Wl,-z,max-page-size=16384) for Android 15+ compatibility.

Debug Logging

The native C++ code includes comprehensive step-by-step debug logging for troubleshooting. Logs are output to:

• Android: Logcat with tag OnnxGenAI (use adb logcat -s OnnxGenAI:*)
• iOS/Desktop: stderr

To disable debug logging in production builds, edit src/flutter_onnxruntime_genai.cpp and set:

#define ONNX_DEBUG_LOG 0

Model Files

ONNX GenAI models must be present on the device before inference. The model directory should contain:

• config.json
• Model weights (.onnx files)
• Tokenizer files

Building from Source

Prerequisites

• Android NDK (set ANDROID_NDK_HOME)
• Xcode command line tools
• CMake 3.20+
• Python 3.8+

Build Commands

# Build for Android only
./scripts/build_onnx_libs.sh android

# Build for iOS only
./scripts/build_onnx_libs.sh ios

# Build for both platforms
./scripts/build_onnx_libs.sh all

# Clean build artifacts
./scripts/build_onnx_libs.sh clean

License

This project is licensed under the MIT License - see the LICENSE file for details.

ONNX Runtime GenAI is licensed under the MIT License by Microsoft.