Why it matters

QLoRA changed FT economics. Understanding shapes strategy.

Advertisement

The architecture

Load base in NF4 4-bit.

Train LoRA in bf16.

Paged optimizers for OOM protection.

QLoRA stackBase NF4 4-bitfrozenLoRA bf16trainablePaged optimizerOOM protectionDettmers 2023; 70B on single H100; PEFT + bitsandbytes standard
QLoRA.
Advertisement

How it works end to end

Dettmers 2023.

NF4 quantization.

Double quantization.

Paged AdamW.