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-{F32,F16,Q8_0,Q5_K_M,Q4_K_M}.gguf	pick one encoding
autoencoder (SAME-L)	stable-audio-3-medium-same-l-v1.0-{F32,F16,Q8_0,Q5_K_M,Q4_K_M}.gguf	must match the DiT
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).

Encodings

sa3-generate --encoding resolves the DiT and the SAME with the same suffix, so download the pair.

encoding	DiT	SAME-L	total
F32	5545 MB	3251 MB	8796 MB
F16	2773 MB	1626 MB	4399 MB
Q8_0	1483 MB	865 MB	2348 MB
Q5_K_M	1053 MB	617 MB	1670 MB
Q4_K_M	962 MB	570 MB	1532 MB

q4_k_m and q5_k_m promote the attention V, feed-forward down and embedding tensors to Q6_K; q8_0 is uniform. Every tier passes sa3-quant-check with below-threshold=0 at cosine 0.990 against the F16 reference, for the DiT and the SAME alike.

Quantization buys footprint everywhere and speed only on some backends. CUDA and Vulkan gain roughly 33% end to end. Metal is flat — Q8_0 is 1.7% faster and Q4_K_M 2.1% slower than F16, because the load-time saving and the added per-step dequant cancel out. On a Mac, pick a quant for the memory, not for the speed.

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

For a quantized set, pass the encoding to both — the downloader and the generator use the same names:

python tools/download_models.py --variant medium --encoding q4_k_m
sa3-generate --model medium --encoding q4_k_m --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.