🎛️ audio.cpp — Low-End GPU Fork

MiniMax Music 3 on a single RTX 3060. Every track below was generated on one 12 GB card.

3.2× faster RTX 3060 12 GB peak VRAM 9 GiB 20 s song in 43 s

Ambsd — “Thirty Sixty”

2:24 · generated in 361 s · 142 BPM · G minor · trap

Ambsd on the build, running on a thirty sixty
Three times faster and the music still pretty
Forty three seconds where it used to be a city
Low end GPU but the output not gritty

Trap demo

2:00 · generated in 293 s · 140 BPM half-time · F minor

Where the 43 seconds goes

Flow matching17.8 s
Autoregressive (LM + depth)17.4 s
Model load4.3 s
Vocoder2.9 s

Source on GitHub · built by @AmbsdOP