GAME-MLX-large

GAME (Generative Adaptive MIDI Extractor) — MLX version of openvpi/GAME GAME-1.0-large checkpoint, for Apple Silicon GPU inference.

File	Description
weights.fp16.safetensors	fp16 weights (low memory)
weights.safetensors	fp32 weights (bit-aligned with torch)
config.yaml	architecture & features config (same as upstream)
lang_map.json	language map (en/ja/yue/zh)

Parameters: 98,988,051 (98.99M) / 1311 tensors. Key names preserved from the torch checkpoint (including model. Lightning prefix).

Usage

pip install gamemlx
gamemlx extract vocals.wav -m large --language zh

import gamemlx
model, lang_map = gamemlx.load_model("large")
gamemlx.extract(model, "vocals.wav", language="zh")

Alignment & performance

Verified on 20 test files against the official PyTorch implementation (CPU, 32-true, identical injected d3pm randomness): fp32 = 20/20 files with all onset/offset/pitch diffs 0.0000; fp16 = 19/20 (pitch diff ≤ 0.0011). MLX fp32 is ~9.4× faster than torch CPU (0.172s vs 1.609s avg per file), peak wired GPU memory 206MB (fp16: 111MB).

Also available on ModelScope: baicai1145/GAME-MLX-large

License

MIT (following upstream openvpi/GAME).

Code: baicai-1145/game-mlx — pip install gamemlx