modelport_executorch

PyTorch ExecuTorch engine for ModelPort. Runs the executorch variants (.pte programs) of modelport.json bundles through executorch_flutter.

await ModelPortFlutter.init(adapters: [ExecuTorchAdapter()]);
final detector = await ObjectDetector.load(
  'https://github.com/ayanparvaiz/modelport/releases/download/zoo-v1/yolos-tiny.json',
);

• The fastest engine on the test phone. MobileNetV3 Small ran in 17 ms on an OPPO CPH1937, against 77 ms with ONNX Runtime, and passed its golden check.
• Small. About 7.6 MB in an arm64 APK.
• Inputs go in manifest order. ExecuTorch takes inputs by position, so the adapter orders them from the manifest and checks every output's type.
• Exports that just work. modelport export --target executorch lowers to XNNPACK and makes example inputs contiguous; a channels_last example input otherwise produces a program that rejects ordinary inputs.

Requirements

• Flutter 3.38 or newer.
• cmake on the build machine, for example brew install cmake. executorch_flutter downloads prebuilt ExecuTorch but builds a small wrapper.
• .pte files exported with the ExecuTorch version executorch_flutter ships (1.5 at the time of writing).

Troubleshooting

macOS build fails with deployment target 11.0. executorch_dart passes 11.0 to CMake, which Xcode 27 rejects, and it can stick in the build cache. Set your app's macOS target to 14.0, delete .dart_tool/hooks_runner/shared/executorch_dart/build/, and build again.

Example

The example app classifies a photo with MobileNetV3, detects objects with YOLOS Tiny, and runs both golden checks.

License

Apache-2.0