QLoRAQLoRA

把模型压到 4-bit 再微调，一张消费级显卡也能练。Quantise the model to 4-bit first, then fine-tune — one consumer GPU is enough.

先把模型量化成 4-bit 冻结住，再只训练少量附加参数。

论文实测：单张 48GB 显卡就能微调 650 亿参数的模型。

The model is quantised to 4-bit and frozen; only a small set of added parameters is trained.

Measured in the paper: a single 48GB GPU can fine-tune a 65-billion-parameter model.

它把「微调大模型」从机房拉到了个人电脑。想自己练模型，先看它够不够你的卡。It moved "fine-tuning a large model" from the datacentre to the personal computer. If you want to train a model yourself, check it against your card first.

去 Models 图谱：看这个模型有哪些量化版、按你的显存该下哪一档Models atlas: which quantisations exist and which tier fits your VRAM ↗

回到名词地图看第5站的其他 29 个词 →Back to the map: the other 29 terms in station 5 →

• QLoRA 论文QLoRA paper ↗