Community: 2.9x faster consumer-GPU inference (sub-realtime on RTX 4090) + two model-card findings

Esamina le indicazioni low-VRAM della model card e la frequenza di campionamento dichiarata di 32kHz confrontandole con le misurazioni riportate della diffusers pipeline. Leggi il README del repository per verificare i risultati di ottimizzazione documentati e gli esiti negativi. Il lavoro è completato quando la model card riflette accuratamente le indicazioni di offload supportate e la frequenza di output riportata dalla pipeline, e il contesto delle prestazioni è documentato dove appropriato.

Hi MiniMax team — first, thank you for open-weighting Music 3; it is a
remarkable model. We spent a focused engineering session getting it to run as
fast as possible on a single consumer GPU (RTX 4090, Windows, the diffusers
modular pipeline) and open-sourced the result as a local Suno-style studio:

https://github.com/TheDutchRuler/minimax-music3-studio

Measured results (20s songs, warm, seed-fixed, bf16 reference precision):

The ensemble idea may interest you most: the AR stage is memory-bandwidth
bound (~23GB of weight reads per frame), so K same-prompt variations decoded
in one batch-2K pass amortize the read — the third song is nearly free. All
math stays row-independent and distribution-identical to the reference
sampler (Gumbel-max equivalence unit-tested).

Two findings you may want to reflect in the model card:

1. The low-VRAM snippet is counter-productive for the AR stage. The card
   suggests apply_group_offloading(pipe.language_model, ..., use_stream=True)
   for small cards. Because the Global LLM decodes autoregressively at 25
   forwards/second, per-layer offload re-streams the full 16.4GB across PCIe
   every frame — we measured 10% GPU utilization and effectively no progress.
   Additionally, use_stream=True pins host memory and roughly doubled
   process RSS (31-38GB) on a 61GB machine. Whole-component offload
   (ComponentsManager.enable_auto_cpu_offload() alone) works well.

2. Sample-rate mismatch: the card says 32kHz output, but the diffusers
   pipeline reports and produces 44.1kHz (pipe.sampling_rate == 44100).

Also documented in the repo README: negative results (FP8 weight-only via
torchao measured 2.1x slower on Windows/torch 2.11; per-layer offload above)
so others don't repeat them.

This work was engineered end-to-end with Claude (Fable 5 Max) by Anthropic.
Happy to provide more detail on any measurement.