Apple Silicon MLX version proof of concept
I got this to work on Apple Silicon M5Pro 64GB. Likely, it will work on all the M series processor with enough RAM. It can currently only be used reliably currently to generate 30s of audio, there are some bug with longer sequence (weird audio effect) and you need 128GB RAM to run it properly (tries to allocate a ton of RAM, like 76Gi for DIT and vocoder stages). 64GB of RAM can successfully generate the AR part for 4500 frames though. It seems that in use, this will use 50GB system RAM during 30s generation.
For 30s generation at least, q4 is usable and increases the speed of generation significantly.
https://huggingface.co/liminalsunset/minimax-music3-mlx
Here's an example generated using the 'quant' branch, using a prompt and lyrics from the official demo website and seed=999.
Disclosure: the repo and the code are entirely generated using LLM and the quality or correctness is not verified.