Fluttorch

Quickstart

Export a model, generate its API, and run the gate.

Export

import torch
from fluttorch_export.export import export_model


class TwoLayer(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = torch.nn.Linear(4, 8)
        self.act = torch.nn.ReLU()
        self.fc2 = torch.nn.Linear(8, 3)

    def forward(self, x):
        return self.fc2(self.act(self.fc1(x)))


model = TwoLayer().eval()
example = torch.randn(1, 4)

export_model(
    model=model,
    example_inputs=(example,),
    name="two_layer",
    out_dir="build",
    input_names=["features"],
    output_names=["score"],
    # References captured from the source model, before lowering. Without these
    # the bundle is a smoke test rather than coverage, and the exporter says so
    # rather than inventing a distribution it has no way to know.
    golden_inputs=[(torch.randn(1, 4),) for _ in range(4)],
)

Three files come out together: two_layer.pte, two_layer.fluttorch.json, and a goldens/ directory. They are one bundle. Moving one without the others is the failure the weight hash exists to catch.

Generate

Put the manifest under lib/ and let build_runner do the rest.

dev_dependencies:
  build_runner: ^2.16.0
  fluttorch_gen: ^0.6.0

dart run build_runner build --delete-conflicting-outputs

You get two_layer.fluttorch.dart: a type per tensor, the spec each was declared with, the manifest carried verbatim so loading needs no separate plumbing, and any preprocessing the export recorded, turned into arithmetic.

final features = TwoLayerFeatures(Float32List.fromList([1, 2, 3, 4]))
  ..preprocess();

The wrapper costs nothing at run time and makes passing the wrong tensor a compile error.

Run the gate

final goldens = await DirectoryGoldenBundle.open('build/two_layer.fluttorch.json');
final model = await ExecuTorchRuntime(bindings).load(
  artifact: await File('build/two_layer.pte').readAsBytes(),
  manifest: goldens.manifest,
);

await expectParity(model, goldens: goldens);

expectParity fails the test. measureParity hands back the report instead, which is what you want when you would rather read the number than react to it:

final reports = await measureParity(model, goldens: goldens);
print(reports.first.describe());

Running this repository’s own suites

The Dart suites need nothing but the SDK. The Python ones need flatc on PATH, which an ExecuTorch checkout builds and nothing puts there for you:

export PATH="$HOME/.cache/fluttorch/executorch/cmake-out-android/third-party/flatc_ep/bin:$PATH"

Without it the export tests fail with FileNotFoundError: 'flatc' from inside a serialization step, which names nothing you would recognise as the cause.

The on-device suite additionally needs the native library, which packages/fluttorch_executorch/tool/build_native.sh builds against an ExecuTorch checkout. tool/prepare_executorch.sh performs the upstream workarounds that checkout needs first. Where the library is absent the suite skips by name rather than silently, so a green run still says what it did not cover.

This site is open source. Improve this page.