VRAM estimates for inference, LoRA, QLoRA, full fine-tuning, and pre-training using fixed task multipliers.
Limited data: VRAM is a catalog estimate times a fixed task multiplier. Quality percents are placeholders, not evals.
In plain English: pick a model and training style, see how much VRAM you need before you buy or rent GPUs.
Compare cloud GPU rental prices to find the most cost-effective setup for your training run.
Also