lib_onnx 0.3.0

ONNX Runtime inference for Dart and Flutter with explicit execution policies and experimental CoreML and QNN integration.

Changelog

0.3.0 - 2026-11-12

• The build hook now bundles official ONNX Runtime CPU inputs for mobile targets: Android arm64-v8a and x86_64 from the Maven Central onnxruntime-android 1.31.1 AAR (runtime and bridge ELF files are checked for 16 KB pages), and iOS device and simulator (arm64, x64) from the official onnxruntime-c 1.30.0 CocoaPods archive, statically linked into the bridge. iOS requires iOS 15.1+.
• iOS uses ONNX Runtime 1.30.0 because no official 1.31.1 iOS archive is published; the loader expects 1.30.0 on iOS and 1.31.1 elsewhere. The bridge uses only C API functions available in both versions.
• Targets without a pinned artifact or native_bundle (for example Windows, Android armeabi-v7a, macOS x64) no longer fail the consumer build: the hook emits no code asset, and creating a runtime throws UnsupportedError.
• Archive URLs may be pinned per target; zip-based archives (.zip, .aar) are unpacked with unzip, falling back to tar.
• Bridge 0.3.0, ABI 2 unchanged. QNN/HTP, CoreML hardware observation and Qwen acceptance remain open.

0.2.0 - 2026-11-05

• Stable CPU inference on macOS arm64 with bridge ABI2 and ONNX Runtime 1.31.1. CoreML/QNN, iOS/Android and full Qwen acceptance remain experimental or incomplete.
• Release workflows precompile macOS arm64 and Linux x64/arm64 bundles, retain license notices and binary manifests, and publish source archives with SHA-256 checksums. Linux bundles remain CPU correctness inputs outside the stable platform claim.
• Updated package, example and native bridge versions together; stable publication retains the platform acceptance gate.

0.2.0-dev.1 - 2026-11-04

• Breaking: bridge ABI2 and Dart3.10+ native-assets packaging; pure Dart core no longer depends on Flutter. Old bridges/ORT mismatches are rejected.
• Checked tensor sizes, scalars/empty tensors/arbitrary ranks, complete partial-result cleanup, task-owned input copies and retained result handles.
• Real asynchronous initialization/loading/run/close, per-task termination, bounded scheduling, monotonic timeouts, removable cancellation listeners and independent errors/diagnostics.
• Enforced ORT CPU fallback constraint, explicit unsupported whitelist/hardware verification results, correct CoreML options/content caches, QNN HTP selection independent of Apple hardware.
• Removed inferred hardware observations and simulated example results; model identity includes external data. Added adapter safety/fp16 support, real small-model example and pinned genuine Qwen validation workflow.
• Added sanitizer/lifecycle/clean-consumer checks and release acceptance gates. iOS/Android accelerator and full Qwen acceptance remain open; no stable release or production acceleration claim.

0.1.0

Historical release notes below are retained. The2026-11-04 review found that async, platform readiness,16KB and hardware-observation claims were not established by the original evidence. They are superseded by the remediation report and current device matrix.

• Initial release of lib_onnx: Hardware-accelerated ONNX Runtime inference library for Dart and Flutter.
• Supported execution providers:
  • Apple CoreML Execution Provider (Apple Silicon GPU & Apple Neural Engine / ANE) on macOS (arm64) and iOS.
  • Qualcomm Snapdragon QNN HTP Execution Provider (Hexagon NPU) on Android (arm64-v8a) with 16 KB page compatibility.
  • ORT CPU Execution Provider for numeric baseline and explicit fallback.
• Pure C ABI bridge with opaque handles, non-blocking native worker thread pool for Flutter UI isolate safety.
• Three-dimensional ExecutionPolicy: Provider priority list, CPU fallback policy (strict, allowed, whitelistedOps), and hardware requirement (any, gpu, npu).
• Transparent 6-level hardware evidence system (DiagnosticReport and EvidenceLevel).
• Numerical tensors: float32, float16, int64, int32, int8, uint8, bool.
• Cooperative per-request cancellation (CancellationToken).
• EmbeddingAdapter implementing official Qwen3-Embedding Last-token pooling and L2 normalization.
• Interactive Flutter example app in example/.
• Automated test suite covering tensor operations, execution policy, embedding adapter, lifecycle, and 1,000 continuous inferences.