gamemlx

Pure MLX inference for GAME (Guided Automatic Music Extraction) — vocal → MIDI, running on Apple Silicon GPU. Numerically aligned with the torch reference; fp32 output is bit-comparable on the test set (20/20 files, all onset/offset/pitch diffs 0.0000) and ~9.4× faster than torch CPU.

Install

uv tool install gamemlx   # or: pip install gamemlx

Quick start

# download weights (HF: baicai1145/GAME-MLX-medium, ModelScope fallback) and extract
gamemlx extract vocals.wav -m medium --language zh

# local model dir, fp16, custom decoding params
gamemlx extract song/ -m ./weights/GAME-1.0-medium-mlx \
    --precision fp16 --language zh --tempo 120 \
    --seg-threshold 0.2 --seg-radius 0.02 --est-threshold 0.2 \
    --t0 0 --nsteps 8 --output-dir out/

Python API:

from gamemlx import load_model, extract

model, lang_map = load_model("medium", precision="fp32")
mids = extract(model, "vocals.wav", language="zh", rng_seed=42)

Models

Weights convert 1:1 from official GAME checkpoints (torch state_dict keys kept verbatim). First use downloads to ~/.cache/gamemlx; ModelScope mirrors are used automatically when HF is unreachable.

Alignment & performance (vs GAME torch CPU, 20-file set)

Details: docs/ALIGNMENT.md, docs/BENCH.md.

Notes

• d3pm sampling randomness is injectable (rng_matrices / --rng-seed) for reproducible comparisons against torch.
• fp32 recommended for exact parity; fp16 saves memory (~111MB wired) with sub-0.002 pitch deviation.