stable-audio-3-medium-base-GGUF

Stable Audio 3 Medium Base DiT — GGUF for sa3.cpp

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

This is a training-only base DiT. Train the adapter on this model, then apply that adapter to Stable Audio 3 Medium for inference. Do not use the post-trained/ARC medium DiT as the training base.

Files

The tokenizer, T5Gemma encoder, conditioner, SAME autoencoder, and inference DiT are fetched from the normal medium model set; they are intentionally not duplicated here.

python tools/download_models.py --variant medium --encoding f16 --training-base
build-cuda/bin/sa3-train --model medium --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 — steady state, medium, 128 frames, dora-rows:

A 2000-step DoRA trained on the Q4_K_M base is audibly indistinguishable from the same run on the F16 base; loss tracks 2.5-4% higher throughout, the offset a slightly degraded forward should produce. Peak VRAM drops from 5243 to 3414 MiB, which is the part that decides whether a small machine can train at all.

python tools/download_models.py --variant medium --encoding q4_k_m --training-base
build-cuda/bin/sa3-train --model medium --models-dir models \
  --dit models/stable-audio-3-medium-base-dit-1.5B-v1.0-Q4_K_M.gguf \
  --dataset /path/to/dataset --adapter-type dora-rows --rank 16 --out train-runs/example

The resulting adapter is an ordinary GGUF LoRA — nothing about it is quantization-specific, so it applies to an F16 or a quantized inference DiT either way.

Provenance

• Source revision: b32993f73c3bdc3864043a72d8032606bba737c8
• Conversion: tools/convert_dit.py --variant medium --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. Organizations above the license's revenue threshold must obtain the appropriate license from Stability AI.