thepatch/stable-audio-3-medium-GGUF

Stable Audio 3 Medium — GGUF (for sa3.cpp)

GGUF conversions of stabilityai/stable-audio-3-medium for sa3.cpp — a portable C++/GGML port of Stable Audio 3, no PyTorch in the loop. Runs on CPU, CUDA, Vulkan, or Metal (Apple Silicon). Every component is validated against the PyTorch reference at cosine similarity ~1.0.

Files

This is a multi-file model. Grab the DiT + SAME at your chosen precision and the conditioner, plus the shared encoder + tokenizer from the t5gemma-b-b-ul2-GGUF repo.

component	file	notes
DiT (diffusion transformer)	stable-audio-3-medium-dit-1.5B-v1.0-{F16,F32}.gguf	pick one precision
autoencoder (SAME-L)	stable-audio-3-medium-same-l-v1.0-{F16,F32}.gguf	pick one precision
conditioner	stable-audio-3-medium-conditioner-v1.0-F32.gguf	tiny sidecar (prompt padding + seconds_total)
encoder + tokenizer	→ t5gemma-b-b-ul2-GGUF	shared across all SA3 variants

F16 is the production path (~3.5s for 12s of audio on an 8GB laptop GPU); F32 is for CPU validation. The conditioner + encoder + tokenizer stay F32 (small / quality-critical).

Usage

For use with sa3.cpp:

# pip install huggingface_hub
python tools/download_models.py --variant medium --encoding f16   # fetches this set + the shared encoder

# --model resolves the 5 gguf files in ./models by name
sa3-generate --model medium --prompt "upbeat funk groove with slap bass" --out song.wav

Performance

Roughly 3s for a 12s clip at f16 on an 8GB laptop GPU (RTX 5070), and ~6s on an Apple M4 — end to end, including model load. The sliding-window decoder keeps long generations linear (a 2-minute clip is ~9s on the 5070). CPU works but is ~10× slower. Full numbers + the f16 / flash-attention levers: docs/BENCHMARKS.md.

License

These are format conversions of stabilityai/stable-audio-3-medium, whose weights Stability AI releases under the Stability AI Community License: free for organizations under $1M annual revenue, with commercial use, fine-tuning, and derivative works permitted within that threshold (above it, contact Stability AI for an Enterprise License). Outputs are yours. That license carries over to these converted weights.

The upstream stable-audio-3 source code is released separately under MIT. Pair these with the shared T5Gemma text encoder, which is Google's under the Gemma Terms of Use.

Relationship to the original

Format conversions (weights → GGUF) for inference in sa3.cpp — no retraining, no architectural changes. See sa3.cpp/docs/DISTRIBUTION.md for the naming convention and how the pieces fit together.