Tim Dettmers

Rating

Domains

Scouting report

Worked out how to run and fine-tune large models in 8-bit and then 4-bit without wrecking them, shipped it as bitsandbytes, and followed with QLoRA — which put fine-tuning a 65B model on a single consumer GPU. His hardware guides are the reference for anyone deciding what to buy.

Curator’s note

Turned "you need a datacentre" into "you need one good GPU".

Portrait

Painted from a photograph of Tim Dettmers by Tim Dettmers, supplied by the subject.

Sources

• 01QLoRA: Efficient Finetuning of Quantized LLMs (2023) arxiv.org
• 02bitsandbytes github.com