this is where we're keeping the version history that used to live at the top of the main README. the README should stay focused on what gary4local is now; this file gets to remember how we got here.
v0.2.0 swaps the SA3 LoRA trainer over to stable-audio-3's own Lightning
trainer instead of the hand-written training loop we were carrying. it's the
same code the inference service already uses, so the training step matches the
reference exactly instead of slowly drifting from it. pytorch-lightning
installs itself the first time you train, so you shouldn't need rebuild env
for it.
the shared trigger word finally does what you'd expect. it gets prepended to
every caption, so you end up training on my-trigger, some caption and the word
comes to mean the style. before this it was replacing the caption around 60%
of the time, which meant a lot of steps trained on a bare trigger with nothing
describing the audio. if you trained a LoRA with a trigger word on an older
version, it's worth retraining it. the "prepend shared phrase" toggle in the
prompt editor is gone too — sidecars are just captions now, and the trigger is
handled for you.
two LoRA annoyances are fixed:
- a freshly trained LoRA didn't show up in gary4juce until you opened the add loras popup. the trainer wrote the catalog, but nothing rebuilt the registry until that window happened to open.
- switching a LoRA's training checkpoint while the SA3 service was running silently did nothing, because adapters get baked in when the model loads. it now reloads the service so the switch actually takes effect.
training should also sit better on 8 GB cards. the VAE and the T5Gemma text encoder get freed or pushed to CPU once they aren't needed, which is roughly 2 GB back, and the logs now show the caption for each clip plus the trigger word being applied so you can see what you're actually training on.
compatible with gary4juce v4.0.7.
v0.1.19 adds full auto-labeling to the SA3 trainer. it uses the ACE-Step captioner to pull genre keywords for SA3 prompt sidecars, alongside the local BPM and key helpers. use rebuild env before using the new SA3 trainer features.
both integrated trainers are more resilient now: caption/preprocessing and checkpoint checks are stricter, interrupted launchers and child workers recover cleanly, and Windows now owns managed inference and training processes as a group so an unexpected app exit cannot leave GPU workers running. closing to the tray still leaves work running as intended.
compatible with gary4juce v4.0.6.
v0.1.18 fixes Carey seed reporting so the seed shown after a random generation is the seed that actually produced the audio.
ACE-Step LoRA sidecars now treat lyrics as BYOL metadata. the captioner still helps with captions, genres, BPM, and key metadata, but it no longer writes LM-hallucinated lyrics into vocal sidecars. the sidecar editor shows a grey lyrics template instead, so users can paste or write the real words themselves.
compatible with gary4juce v4.0.4.
v0.1.17 adds carey seed support for lego, complete, and cover, plus an optional ScragVAE decoder toggle for ACE-Step.
it also keeps the local ACE-Step LoRA trainer moving forward. early testing says
regular acestep-v15-base is still the safer lego model when no LoRA is loaded,
while xl-base gets much more exciting once you have a matching xl-base LoRA,
especially for vocals and backing vocals.
this is a small hotfix for carey environment rebuilds on clean machines.
- pin
trove-classifiers==2026.5.22.10for Carey's isolated Hatchling build environment because the newer2026.6.1.19wheel is missing required package-name metadata - treat omitted and explicit
nullsampling controls the same way in gary - initialize Stable Audio's model duration before validating explicit loop bars
- clarify that carey's completion duration means final total duration, not seconds appended
compatible with gary4juce v4.0.2.
we've got integrated ACE-Step LoRA training now.
- caption and prepare ACE-Step datasets with the 0.6B, 1.7B, or 4B captioner
- edit captions, genres, BPM, key, lyrics, and other sidecar metadata before training
- train LoRA or DoRA adapters against regular base or XL-base
- use Min-SNR loss weighting, best-checkpoint tracking, and the experimental balanced attention + MLP profile
- offload frozen model components and run a conservative VRAM preflight before the first batch
- automatically repair safe missing captioning/training dependencies
- keep standard and XL LoRAs and caption pools isolated automatically
- register completed adapters and their prompt pools with carey
the carey service also handles model offloading more cleanly now, which makes it much easier to swap between base, turbo, SFT, and XL models while using gary4juce.
see the ACE-Step LoRA training guide for the honest version of what we've tested and what remains experimental.
fair warning... this trainer has only been tested on a 5070 laptop GPU with
training runs using ace-step-v15-base. plz let me know if you have any issues
with xl-base.
v0.1.14 makes Hugging Face permission failures explicit when downloading gated Stable Audio 3 models.
- preserves the underlying Hugging Face error when a generic cache error wraps
a
401or403response - explains when a fine-grained token needs public gated-repository read access
- labels a stored token as saved rather than implying its permissions have already been validated
- places the gated-token permission guide directly on the sa3 model screen
- repairs older sa3 environments by installing missing LoRA training dependencies before preprocessing begins
- cancels sa3 LoRA training without flashing PowerShell or taskkill windows
v0.1.13 hardens Hugging Face onboarding and model downloads, especially for users who are new to gated repositories and fine-grained access tokens.
- adds an in-app visual guide for enabling public gated-repository access on a fine-grained Hugging Face token
- shows actionable model-download errors directly in the model list
- uses Hugging Face's official snapshot downloader for resumable downloads and reliable cache layout on Windows
- detects incomplete sa3 snapshots instead of presenting them as ready
- loads complete Stable Audio 3 Medium and bundled T5Gemma files directly from the local cache, avoiding unnecessary Hub checks during inference
v0.1.12 adds Stable Audio 3 LoRA training directly to the Windows control center. the trainer is a focused integration of dada-bots' underfit project, adapted to use gary4local's existing sa3 environment, saved Hugging Face token, model storage, and LoRA registry.
- choose an audio dataset, edit optional text-sidecar prompts, and start training with practical defaults for consumer NVIDIA GPUs
- follow selectable, auto-scrolling logs and persisted job progress, or cancel preprocessing and training from the same window
- copy completed
.safetensorsadapters into gary4local's sa3 LoRA folder and register them for generation automatically - optionally normalize each track's encoded latent RMS to the base-model target before training
- align sampler latents with the decoder's device and precision before decode
- convert half-precision models before moving them to the GPU to reduce peak loading memory