Troubleshooting¶

"no registered adapter can run …"¶

Add the adapter package the error names and pass it to ModelPortFlutter.init(adapters: [...]).

"… needs at least N MB of RAM"¶

Every variant needs more memory than the device reports. Pick a smaller model, or ship a smaller variant first in the manifest.

"sha256 of … does not match the manifest"¶

The file on the server changed after the manifest was written, or the download was corrupted. Republish the bundle with modelport pack and modelport publish.

Android: llama.cpp finds no CPU backend¶

Add packaging { jniLibs { useLegacyPackaging = true } } to android/app/build.gradle.kts. modelport_llamacpp also works around an llm_llamacpp 0.7.0 issue where its model loader does not load backends on Android.

Android: release build crashes in ONNX Runtime¶

Add -keep class ai.onnxruntime.** { *; } to android/app/proguard-rules.pro.

macOS: deployment target errors¶

flutter_onnxruntime needs macOS 14. Xcode 27 also rejects targets below 12. Set MACOSX_DEPLOYMENT_TARGET = 14.0 in macos/Runner.xcodeproj.

macOS: executorch_dart fails with target 11.0¶

executorch_dart passes CMAKE_OSX_DEPLOYMENT_TARGET=11.0 to CMake, which Xcode 27 rejects, and the value can stick in its build cache. Delete .dart_tool/hooks_runner/shared/executorch_dart/build/ in your project or workspace and build again.

ExecuTorch: build fails with "cmake not found"¶

executorch_flutter downloads prebuilt ExecuTorch but builds a small wrapper with CMake. Install it, for example brew install cmake.

ExecuTorch: "Propagating input shapes failed"¶

The .pte was exported from a channels_last example input. modelport export always makes example inputs contiguous; if you export by hand, call .contiguous() first.

Tests: models disappear between Android integration test runs¶

flutter test uninstalls the app at the end, which deletes its data. Pass --no-uninstall.

Everything is slow in tests¶

Integration tests run in debug mode. Use flutter drive --profile to measure.