Skip to content

Comment on QLoRA: Efficient Finetuning of Quantized LLMsparent

Comments

This work introduces a new quantization scheme, NF4, for 4-bit NormalFloat, based on previous work on quantile quantization, so it's not a simple truncation, but it's also not a GPTQ-like optimization method. Figure 3 of the paper shows accuracy improvement of NF4 over FP4.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.