Laya Tetris example

Path: example/laya_tetris · Platforms: macOS 14.0+, iOS 16.4+, Android 10+ (API 29), Web · First-run download: laya-Q8_0.gguf (421 MB), laya-head.safetensors (106 MB) and laya-head-tetris.safetensors (106 MB)

A Flutter app in which a Laya decision model plays real-time Tetris while gravity keeps running. Live demo: https://leehack-flutter-laya-tetris.static.hf.space

Run#

cd example/laya_tetris
flutter pub get
flutter run -d macos   # or an iOS or Android device

Variants, from example/laya_tetris:

# Web: fetch the pinned bridge assets, then run
WEBGPU_BRIDGE_OUT_DIR="$PWD/web/webgpu_bridge" \
  ../../scripts/fetch_webgpu_bridge_assets.sh
flutter run -d chrome

# Headless games from local model files, no downloads
dart run bin/bench.dart --model laya-Q8_0.gguf --head laya-head.safetensors \
  --tuned-head laya-head-tetris.safetensors

# Use your own tuned head
flutter run -d macos --dart-define=LAYA_TUNED_HEAD_URL=<url>

The app caches the downloads in a laya/ folder and reuses them on later launches. A Web build needs a cross-origin isolated page.

What it demonstrates#

• One LlamaEngine holding the backbone GGUF, shared by two DecisionEngines: the base head and a Tetris-tuned head (Decision models).
• First-launch downloads with progress through loadModelSource for the backbone and modelDownloadManager.ensureModel for the heads (Downloads and cache).
• Yes/no and choice questions sent as one systemOneBatch call per piece, or per knockout round, while the llama.cpp worker isolate does the work (Batches).
• A typed ChoiceKey.of over the candidate placements, read back with answerOf (Choice values).
• Switching GPU and CPU, backbones and thread counts at runtime by disposing the engine and loading a new one (Backend selection).
• The same app on the llama.cpp WebGPU bridge (Decision models: Web).

Test#

cd example/laya_tetris
flutter test

LAYA_MODEL_DIR=<folder> flutter test test/laya_models_local_test.dart also loads the real backbone and base head from <folder>.

Full options: the players, candidate modes, tuned-head lookup order, Web serving, headless benchmark flags and measured results are in the example README; fine-tuning the head is in training/README.md.