SCAO drops standalone LoRA optimizer

SCAO is interesting, but still firmly in “prove it outside the author’s bench” territory.

• –The pitch targets a real pain point: AdamW can waste early fine-tuning steps, especially when local users are constrained by GPU memory and time.
• –The standalone `scao.py` path is the right adoption move after Hugging Face rejection because it lets PEFT users test without waiting for core library approval.
• –Reported results are promising, including 36.7% VRAM reduction, roughly 627 tokens/sec on TinyStories-1M full fine-tuning, and 25.8% GPT-2 perplexity improvement versus AdamW.
• –The weak signal is community validation: the Reddit thread is tiny and the repo is very new, so developers should treat claims as experimental until independent LoRA benchmarks appear.