#paged-optimizers

Question 1: Select all that apply

QLoRA fine-tunes very large language models on a single GPU. Which statements describe how it works? Select all that apply.

Paged optimizers move optimizer states to CPU RAM automatically when GPU memory runs short1
Weight gradients are computed only for the LoRA parameters, which are kept in BFloat162
The frozen base weights are stored in 4-bit NormalFloat and dequantized to BFloat16 for computation3
The 4-bit base weights also receive gradient updates, and the LoRA adapters add extra capacity4
NormalFloat spaces its levels evenly, because pretrained weights are spread out uniformly5