跳转至

ONNX Runtime Pin

flutter_onnxruntime is pinned to an exact version in pubspec.yaml:

flutter_onnxruntime: 1.7.1   # NOT ^1.7.1

Do not widen this to a caret range without working through the re-check procedure below and testing on a physical arm64 Android device.

Why

v1.0.2 shipped a crash: the app died instantly when the user started a recording. Two Play Console native traces showed a SIGSEGV deep inside libonnxruntime.so, under OrtSession.run, with byte-identical PC offsets — a deterministic kernel fault, not a race.

The crashing library was ONNX Runtime 1.23.0, confirmed by build ID:

ONNX Runtime arm64 libonnxruntime.so build ID
1.22.0 c3b77e8b1f5d988f59094b66325830fe6e6d3af4
1.23.0 34a6804242f0b38bb97f611351b9bc0630af11ae ← in both crash reports

Nothing in the app changed. The native runtime was swapped underneath us:

  • pubspec.yaml declared flutter_onnxruntime: ^1.7.0.
  • pubspec.lock was gitignored, so nothing pinned the resolution.
  • flutter_onnxruntime 1.8.1 bumped the Android native dependency from onnxruntime-android:1.22.0 to 1.23.0.
  • v1.0.1 was built against 1.7.1 (ORT 1.22.0); v1.0.2 resolved 1.8.3 (ORT 1.23.0) roughly a day later.

ORT 1.23.0 routes fp32 GEMM through Arm KleidiAI on arm64, which 1.22.0 did not. Symbols present only in 1.23.0:

  • ArmKleidiAI::MlasGemmBatch
  • SME/SME2 micro-kernels (..._sme2_mopa, f32p2vlx1, kai_lhs_pack_f32p2vlx1_f32_sme)
  • kai_get_sme_vector_length_u32

1.22.0 only carries KleidiAI's int4/int8 dotprod/i8mm kernels — nothing on the fp32 path. The BirdNET classifier is fp32 conv/matmul, so with 1.23.0 every inference enters this new code. The plugin author documented a crash in exactly this path in the 1.8.1 changelog:

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 (upstream microsoft/onnxruntime#29538)

They fixed it for iOS by downgrading to 1.23.0, but moved Android up into 1.23.0 in the same release.

There is no runtime escape hatch: ORT exposes no session-config key to disable KleidiAI (the binary only defines session.disable_cpu_ep_fallback, session.intra_op.allow_spinning, session.intra_op_thread_affinities). The version pin is the only lever.

Re-check procedure

Revisit when either of these lands:

  • upstream microsoft/onnxruntime#29538 is closed/fixed, or
  • flutter_onnxruntime ships a release whose changelog moves Android to an ORT version newer than 1.23.0 that includes the KleidiAI fix.

How to re-check:

  1. Bump pubspec.yaml to the candidate version and flutter pub get.
  2. Confirm which native runtime you actually got: 
    grep onnxruntime-android \
  ~/AppData/Local/Pub/Cache/hosted/pub.dev/flutter_onnxruntime-<ver>/android/build.gradle
  3. Check whether the fp32 KleidiAI path is still compiled in. Extract the AAR from ~/.gradle/caches/modules-2/files-2.1/com.microsoft.onnxruntime/ and scan jni/arm64-v8a/libonnxruntime.so for ArmKleidiAI, _sme2_mopa, and f32p2vlx1. Their presence is not automatically a blocker once upstream is fixed — but it means step 4 is mandatory.
  4. Run a real Live session on a physical arm64 device for several minutes. The crash reproduces on the first inference, so a session that survives a few minutes of continuous detection is a strong signal. An emulator is not a valid test: it may not expose the SME/SME2 CPU features that select the affected kernels.
  5. Verify the shipped build ID matches what you tested before releasing.

If the candidate is clean, update this document with the new evidence rather than deleting it.

pubspec.lock is now tracked (see the !pubspec.lock negation in .gitignore). This app is an application, not a library — releases must be reproducible, and a native runtime swap must show up as a reviewable diff. Always commit pubspec.lock changes deliberately, and re-read this document when the diff touches flutter_onnxruntime.