v0.2.0 swaps the SA3 LoRA trainer over to stable-audio-3's own Lightning trainer instead of the hand-written loop we were carrying. it's the same code the inference service already uses, so training matches the reference instead of drifting from it. pytorch-lightning installs itself the first time you train, so you shouldn't need rebuild env for it.
you can now train on full tracks instead of random crops. the new toggle uses the model's real native length (~285s) so each clip trains whole, which is what our best LoRAs were trained on. it's on by default and does use more VRAM than short crops, but with the new VAE and T5Gemma freeing it's still quick on my 5070 laptop GPU. not painful at all, and worth it for me.
the shared trigger word now gets prepended to every caption instead of replacing it, so my-trigger, some caption teaches the word to mean the style. before, 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. worth retraining any trigger LoRA you made on an older version.
two LoRA fixes: a freshly trained LoRA no longer stays invisible until you open the add loras popup, and switching a LoRA's training checkpoint while the SA3 service is running now actually takes effect.
training also fits better on 8 GB cards. the VAE and T5Gemma encoder get freed or pushed to CPU once they aren't needed, which is roughly 2 GB back.
compatible with gary4juce v4.0.7.