Skip to content

Comment on QLoRA: Efficient Finetuning of Quantized LLMs

Comments

Do you know which model size can be run with a 3090?

3090 can handle the ~30B models quantized to 4 bits.

Assuming you have 24GB of VRAM, then you should be able to run something like 7 billion parameters i.e. MPT-7b and quantized perhaps you can get up to 13 billion.

Note: I have no practical experience of this, just reading around.

24GB can fit 33B parameter models in 4bit. You only need 4GB to run 7B models.

AboutSource Built by g1lg1l

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