resolveHuggingFace static method

Future<ResolvedHfModel> resolveHuggingFace(String repo, { ModelFileType? fileType, String? token, PreferredBackend? preferredBackend, })

Resolves a Hugging Face repo id into a ResolvedHfModel by reading that repo's deployment metadata (e.g. litertlm_manifest.json), using a resolver registered via initialize — either passed in huggingFaceResolvers: or auto-derived from a registered engine that ships one (LiteRtLmEngine, OnnxEngine, BuiltInAiEngine).

This is the INSPECT-ONLY path (resolve, then install yourself). To resolve AND install in one call, use installModel(...).fromHuggingFace(repo) with no file: — it resolves the manifest internally, installs the revision-pinned variant, and returns the runtime defaults on InferenceInstallation.runtime.

fileType tells the registry which resolver to pick (the litertlm resolver claims ModelFileType.litertlm); it is the only deterministic selection signal a resolver has before it fetches the manifest.

The result carries install identity (ResolvedHfModel.modelType / ResolvedHfModel.fileType for installModel, and ResolvedHfModel.url for the download) plus overridable ResolvedHfModel.runtime defaults. Apply the model-level defaults with getActiveModel(defaults:); forward the two session-level fields to createSession yourself. Install from ResolvedHfModel.url (not fromHuggingFace(repo, file:)) so a resolver's pinned revision is honoured — the plugin still auto-applies the HF token to huggingface.co hosts:

final r = await FlutterGemma.resolveHuggingFace(
    repo, fileType: ModelFileType.litertlm);
await FlutterGemma.installModel(
      modelType: r.modelType ?? ModelType.general,
      fileType: r.fileType,
    )
    .fromNetwork(r.url)
    .install();
final model = await FlutterGemma.getActiveModel(defaults: r.runtime);
// minOutputTokens is a FLOOR, not a cap — leave maxOutputTokens unset (or
// keep it >= r.runtime.minOutputTokens); passing the floor AS the cap
// would truncate a reasoning model mid-thought.
final session = await model.createSession(
  enableThinking: r.runtime.isThinking ?? false,
);

token defaults to the token passed to initialize. Throws StateError if no registered resolver handles repo — add the engine package that provides one (e.g. flutter_gemma_litertlm for .litertlm manifests).

Implementation

static Future<ResolvedHfModel> resolveHuggingFace(
  String repo, {
  ModelFileType? fileType,
  String? token,
  PreferredBackend? preferredBackend,
}) async {
  final resolver = HuggingFaceResolverRegistry.instance.findFor(
    repo,
    fileType: fileType,
  );
  if (resolver == null) {
    throw StateError(
      'No Hugging Face resolver registered for "$repo" (fileType: $fileType). '
      'Register the engine that provides one — its resolver auto-registers via '
      'HuggingFaceResolverSource (e.g. LiteRtLmEngine ships the litertlm '
      'manifest resolver) — or pass one explicitly to '
      'FlutterGemma.initialize(huggingFaceResolvers: [...]).',
    );
  }
  return resolver.resolve(
    repo,
    token: token ?? ServiceRegistry.instance.huggingFaceToken,
    platform: _currentPlatformKey(),
    preferredBackend: preferredBackend,
  );
}