FFT

Claim: FFT halves LLM finetuning VRAM — real optimization or just hype?

Reddit dev claims swapping AdamW for FFT halves LLM finetuning VRAM, but slows training. Unverified claim from Spectra-Global awaits peer review.