LoRA：LoRA: Low-Rank Adaptation of Large Language Models

ICLR 2022 · Microsoft

冻结原权重，只训练低秩分解的增量矩阵，参数量降低上万倍，推理时可合并、零额外延迟。