PEFT — Parameter-Efficient Fine-Tuning

PEFT is the Hugging Face library for fine-tuning large language models without updating all parameters. Using techniques like LoRA, QLoRA, and IA3, you can adapt a 7B+ parameter model on a single consumer GPU in hours while matching or exceeding full fine-tuning quality on downstream tasks.

Key features

• LoRA — inject low-rank adapter matrices; only train ~0.1% of parameters
• QLoRA — quantize base model to 4-bit, then add LoRA adapters — fine-tune 70B models on 48 GB VRAM
• Prompt tuning & prefix tuning — learn a soft prompt prefix instead of touching weights
• IA3 — learns rescaling vectors; even more parameter-efficient than LoRA
• Merge & unload — merge adapters back into base weights for zero-latency inference
• Seamless HF ecosystem — works with transformers, trl, accelerate, and bitsandbytes

Quick start

npx ai-supply add peft-parameter-efficient-finetuning

# Or install directly
pip install peft transformers torch

from transformers import AutoModelForCausalLM
from peft import LoraConfig, get_peft_model

base_model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    lora_dropout=0.05,
    task_type="CAUSAL_LM"
)

model = get_peft_model(base_model, lora_config)
model.print_trainable_parameters()
# trainable params: 4,194,304 || all params: 3,756,523,520 || trainable%: 0.1117

Curated mirror of the open-source PEFT project (Apache-2.0). Install upstream from the repository.