thepatch/stable-audio-3-small-sfx-base-GGUF

Stable Audio 3 Small SFX Base DiT — GGUF for sa3.cpp

F32, F16 and Q4_K_M GGUF conversions of the training DiT from stabilityai/stable-audio-3-small-sfx-base, for native LoRA/DoRA training with sa3.cpp.

This is a training-only base DiT. Train the adapter here, then apply it to Stable Audio 3 Small SFX for inference.

Files

file	size	purpose
stable-audio-3-small-sfx-base-dit-0.5B-v1.0-Q4_K_M.gguf	302 MB	smallest, and the fastest to train on
stable-audio-3-small-sfx-base-dit-0.5B-v1.0-F16.gguf	877 MB	the reference training base
stable-audio-3-small-sfx-base-dit-0.5B-v1.0-F32.gguf	1751 MB	CPU/reference validation
SHA256SUMS		release checksums

The normal small-SFX model set supplies all other components.

python tools/download_models.py --variant small-sfx --encoding f16 --training-base
build-cuda/bin/sa3-train --model small-sfx --models-dir models --dataset /path/to/dataset --out train-runs/example

Training on the Q4_K_M base

Training against a quantized base is supported on every backend — CPU, CUDA, Vulkan and Metal. The frozen base only ever enters the adapter path as a mul_mat argument, and the one backward that needed it, out_prod(W, transpose(grad)), now accepts a quantized src0 on all three GPU backends. It is faster than F16 rather than a tradeoff, on a 2.9x smaller file, and the adapter it produces is an ordinary GGUF LoRA that applies to an F16 or a quantized inference DiT either way.

At 302 MB this is the smallest trainable SA3 base published, which is the one that decides whether a phone or a low-VRAM laptop can train at all.

python tools/download_models.py --variant small-sfx --encoding q4_k_m --training-base
build-cuda/bin/sa3-train --model small-sfx --models-dir models \
  --dit models/stable-audio-3-small-sfx-base-dit-0.5B-v1.0-Q4_K_M.gguf \
  --dataset /path/to/dataset --adapter-type dora-rows --rank 16 --out train-runs/example

Provenance

• Source revision: cc5ddb990e30daa68336ac61c140c37c7033ab7c
• Conversion: tools/convert_dit.py --variant small-sfx --training-base, then tools/quantize_gguf.py for F16 and sa3-quantize --mix q4_k_m for the quant
• Relationship: tensor rename/serialization and precision conversion only; no retraining
• The Q4_K_M base passes sa3-quant-check against F16 with below-threshold=0 at cosine 0.990

License and attribution

Powered by Stability AI.

These converted weights remain under the Stability AI Community License. This repository includes the upstream LICENSE.md and a NOTICE describing the conversion and retaining the required Stability AI attribution.