PEFT — файн-тюнинг LLM через LoRA и QLoRA

★ 7.7 · ai-tooling

peft is a Claude Code skill that implements parameter-efficient fine-tuning of LLMs using LoRA, QLoRA, and 25+ adapter methods from HuggingFace's official PEFT library. It enables fine-tuning of 7B–70B models on consumer GPUs such as the RTX 4090 or A100 by training less than 1% of parameters — resulting in 6 MB adapters instead of 14 GB full-model weights. QLoRA mode combines 4-bit quantization via bitsandbytes with LoRA, fitting a 70B model onto a single 24 GB GPU at roughly a 5% quality trade-off versus full fine-tuning. Dependencies are peft ≥ 0.13.0, transformers ≥ 4.45.0, torch ≥ 2.0.0, and bitsandbytes ≥ 0.43.0. The skill suits teams that need rapid iteration across multiple task-specific adapters sharing one base model under tight memory constraints.