flutter_onnxruntime

What's Changed

• fix: HardSwish returns all zeros on Linux under comma-decimal locales in #74 (reported by @@LeonardoGamar)

Full Changelog: v1.8.3...v1.8.4

• Fix HardSwish and other function-defined ONNX operators returning all zeros on Linux under comma-decimal system locales such as German and French (#73, upstream onnx/onnx#8111)

Changes

• Fix App Store Connect validation errors (ITMS-90208) on iOS/macOS by repackaging ONNX Runtime as a library-format xcframework instead of framework-format (issue #71)

Full Changelog: v1.8.2...v1.8.3

• Fix App Store Connect validation errors (ITMS-90208) on iOS/macOS by repackaging ONNX Runtime as a library-format xcframework instead of framework-format (issue #71)

What's Changed

• Modernize Android build config and fix Kotlin 2.4 compatibility warnings in #70

Full Changelog: v1.8.1...v1.8.2

• Migrate Kotlin options to a newer format and resolve KTLC warnings to make the plugin ready for Kotlin 2.4 (#70)

What's Changed

• Fix memory surge/OOM crash during inference on SME-capable ARM64 devices (iPhone A18+/M4+) caused by a KleidiAI convolution memory regression in ONNX Runtime 1.24.x (#66, upstream microsoft/onnxruntime#29538) (#68). Thanks to @runoob-coder for reporting this and helping debug!
• Move ONNX Runtime to 1.23.0 on all platforms: Android upgraded from 1.22.0; iOS/macOS downgraded from 1.24.2 via a masicai fork of the ONNX Runtime Swift package (Microsoft's SPM repo has no 1.23.x tag)
• Add save model extraction from assets to avoid ORT loading corrupted models (#67)
• Add true bool support on iOS/macOS via a C++ bridge (#69)

Full Changelog: v1.8.0...v1.8.1

• Fix memory surge/OOM crash during inference on SME-capable ARM64 devices (iPhone A18+/M4+) caused by a KleidiAI convolution memory regression in ONNX Runtime 1.24.x (#66, upstream microsoft/onnxruntime#29538) (#68)
• Move ONNX Runtime to 1.23.0 on all platforms: Android upgraded from 1.22.0; iOS/macOS downgraded from 1.24.2 via a masicai fork of the ONNX Runtime Swift package (Microsoft's SPM repo has no 1.23.x tag)
• Add save model extraction from assets to avoid ORT loading corrupted models (#67)
• Add true bool support on iOS/macOS via a C++ bridge (#69)

What's Changed

• Add Swift Package Manager support for iOS and macOS in #62
• Move android integration test CI from macos to ubuntu in #63

Full Changelog: v1.7.1...v1.8.0

• Add Swift Package Manager support for iOS and macOS (#62); CocoaPods remains fully supported
• Upgrade onnxruntime-objc from 1.22.0 to 1.24.2 for iOS and macOS (aligned with the official ONNX Runtime Swift package)
• Move Android integration test CI from macOS to Ubuntu

What's Changed

• Prevent startup crash from makeBackgroundTaskQueue forwarding on macOS in #59

Full Changelog: v1.7.0...v1.7.1

• Prevent startup crash from makeBackgroundTaskQueue forwarding on macOS (#58 #59)

What's Changed

• Use BackgroundTaskQueue to avoid UI thread blocking during inference by @timeofjohnny in #56
• Add synchronization to prevent race condition during engine detach by @biendltb in #57

Full Changelog: v1.6.4...v1.7.0

• Run inference on a background thread via Flutter's TaskQueue for Android, iOS, and macOS
• Prevent race condition introduced by running inference on a background thread in the case of disposing the session while inference is running

What's Changed

• Fix uint8 tensor creation defaulting to int8 on Android by @liyuqian in #53
• Fix uint8 and bool tensor type handling across all platforms by @biendltb in #54

Full Changelog: v1.6.3...v1.6.4

• Fix tensor creation for bool and uint8 types
• Add missing tensor conversions for bool and uint8 types across Android, iOS/macOS, and Web
• Add more integration tests to harness the conversions

What's Changed

• Add float16 tensor support for iOS and macOS in #50
• Upgrade onnxruntime-objc to 1.22.0 for macOS in #51

Full Changelog: v1.6.2...v1.6.3

• Add float16 tensor support for iOS and macOS
• Upgrade onnxruntime-objc for macOS to 1.22.0

What's Changed

• Fix memory leak when creating tensors by @biendltb in #47
• Fix race condition in convertTensor for Windows and Linux by @biendltb in #48

Full Changelog: v1.6.1...v1.6.2

• Fix a memory leak issue for tensor creation in Linux and Windows
• Fix thread safe issue for tensor operations in Linux and Windows

• Explicit cleanup for ONNX Runtime resources when applications close in macOS and iOS (#45)

• Optimize the data transfer from native to Dart using typed arrays (#43)
• Setup cmake-format for CMake formatting (#41)
• Add bug report and feature request templates

• Migrate the web plugin from dart:js to dart:js_interop (#40)

• Upgrade ONNX Runtime to version 1.22.0 for 16KB Page Size support in Android (#39)
• Fix Linux build issues in CMake and X11 display server (#37)
• Upgrade Kotlin version to 2.1.0 (#38)

• Fix an issue with input name-value mismatch in Windows (#28 #29)
• Add more integration tests for input name and order consistency
• Add integration tests for int64 input tensor

• Remove unnecessary warnings in CMake for the Linux build
• Improve README and troubleshooting documentation

• Support string tensors in all platforms
• Reinforce structure and behavior consistency between Linux and Windows implementations
• Minor bug fixes and documentation updates

• Support Windows platform 🎉🎉🎉
• Refactor SessionManager in Linux for cleaner architecture
• Add Azure to provider list
• Improve the CMake build stability in Linux

• Support Web platform 🎉🎉🎉
• Add integration tests for Web in both local and CI

• Standardize Execution Provider names across all platforms
• Fix a bug in get_metadata method for Linux

• Improve example documentation

• Returning a multi-dimensional list for tensor data extraction
• Support return a flat list

• Standardize error codes and error messages from native
• Remove auto disposal of input tensors after inference in Kotlin
• Refactor C++ to return standard Platform Exception and impose standard error handling
• Add back the example to package as required by pub.dev; this will increase the package size unnecessarily

• Support running inference with an ONNX model on Android, iOS, Linux, and macOS
• Support ONNX Runtime version 1.21.0