maxgraf/sa3-variations-torch

SA3 Variations — fp16 weights

Float16 conversions of Stability AI's Stable Audio 3 checkpoints, repackaged for the Windows build of SA3 Variations.

These are Stability AI's models. This repo is a redistribution, not a new model, and is not affiliated with or endorsed by Stability AI.

Why this exists

The SA3 Variations Windows app runs inference through PyTorch and needs its weights on disk at install time. The upstream repos are gated, which would mean every user creating a Hugging Face account and a token just to install a desktop app. This repo removes that step.

What was changed

Converted with scripts/convert_weights_fp16.py: every float32/float64 tensor cast to float16. Integer indices, masks and bools are untouched, and the bundled T5Gemma text encoder is left in bfloat16 — bf16 to fp16 narrows exponent range and risks overflow.

This is a storage change, not a quality one: the app loads with model_half=True on CUDA, so these weights were being cast to fp16 at load anyway. Verified by generating identical seeds from the original fp32 model and this conversion:

corr(fp32, fp16) = +1.000000   (3 seeds)

	upstream fp32	here
medium/model.safetensors	9.22 GB	4.3 GB
medium/t5gemma-b-b-ul2/	1.18 GB	unchanged (bf16)

Layout

Each directory is self-contained — checkpoint, config, and its own copy of the text encoder, so the app can load it with a local path and no network access:

medium/
  model_config.json
  model.safetensors
  t5gemma-b-b-ul2/
small-music/
  model_config.json
  model.safetensors
  t5gemma-b-b-ul2/

Licence

Inherits the upstream terms, which you should read before using these files:

• medium/ — stabilityai/stable-audio-3-medium
• small-music/ — stabilityai/stable-audio-3-small-music

The bundled T5Gemma encoder additionally carries Google's Gemma Terms of Use.

Source

App and conversion script: https://github.com/maxgraf96/stable-audio-3