Default LoRA hyperparameters do not match 16-bit performance. 作者发现按照惯例只将 LoRA 用在 query and value attention projection matrices 上并不能 match 全精度微调,将 LoRA 用在所有线性层上才能 match 全精度微调
Experiments
Memory Footprint
QLoRA vs. Standard Finetuning
4-bit NormalFloat yields better performance than 4-bit Floating Point.
k-bit QLORA matches 16-bit full finetuning and 16-bit LoRA performance.
References
Dettmers, Tim, et al. “Qlora: Efficient finetuning of quantized llms.” arXiv preprint arXiv:2305.14314 (2023).
code: https://github.com/artidoro/qlora and https://github.com/TimDettmers/bitsandbytes