File 002 · Local Intelligence
Record 083 · Tooling · 2023
23.83.FQLoRA
University of Washington
“4-bit fine-tuning that made adapter training a homelab sport.”
Fine-tune a 65B on one GPU.
Dettmers et al. combined 4-bit NF4 quantization with LoRA. A single 24GB card could specialize a large model overnight.
Opened large-model fine-tunes on consumer GPUs.
Filed notes
- NF4 4-bit training
- 65B on 24GB
- bitsandbytes stack
SIC
7372
Prepackaged Software
NAICS
511210
Software Publishers
