Problem

The distributed macOS aarch64 binary (voicebox_aarch64.app.tar.gz) silently falls back to PyTorch CPU instead of using MLX, despite the README advertising 4-5x faster inference on Apple Silicon.

Two root causes:

1. platform_detect.py only catches ImportError

Inside a PyInstaller bundle, MLX loads its Metal shader libraries (.metallib) at import time. When those files are missing from the bundle, Python raises OSError — not ImportError. The original code:

try:
    import mlx
    return "mlx"
except ImportError:   # ← OSError slips through here
    return "pytorch"

This causes a silent fallback to PyTorch on every Apple Silicon machine running the binary.

2. collect_data_files misses native libraries

build_binary.py and voicebox-server.spec used --collect-data / collect_data_files for mlx and mlx_audio. This only copies pure-Python files — native .dylib and .metallib files are excluded. PyInstaller needs --collect-all / collect_all + Analysis(binaries=...) to bundle shared libraries correctly.

Fix

1. platform_detect.py — catch (ImportError, OSError, RuntimeError) and import mlx.core (forces native lib loading eagerly so any failure is caught here)
2. build_binary.py — replace --collect-data mlx/mlx_audio with --collect-all
3. voicebox-server.spec — use collect_all() and pass collected binaries to Analysis(binaries=...)

Testing

Reproduced on macOS 15 (Apple Silicon M4 Mac mini):

• Before fix: binary logs Backend: pytorch, generation takes 30+ min for a 5s clip
• After fix (source build with MLX): logs Backend: MLX | GPU: MPS (Apple Silicon), same clip in ~39s (first run) / ~8s (warm)