2nd/10_Wiki/Topic_Programming/AI_and_ML/PEFT (Parameter-Efficient Fine-Tuning).md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-peft-parameter-efficient-fine-tu	PEFT (Parameter-Efficient Fine-Tuning)	10_Wiki/Topics	verified	self	PEFT Parameter-Efficient Fine-Tuning LoRA fine-tuning	none	A	0.95	applied	peft lora qlora fine-tuning llm		2026-05-10	pending	language framework python peft, transformers, bitsandbytes

PEFT (Parameter-Efficient Fine-Tuning)

매 한 줄

"매 frozen base + tiny trainable delta". 매 full fine-tuning 의 ~0.1-1% parameter 만 학습. 2026 standard: LoRA / QLoRA — 매 70B model 도 single 24GB GPU 에서 fine-tune 가능. HuggingFace peft library 의 사실상 표준.

매 핵심

매 동기

• Full FT: 70B model = 280GB (fp32) gradient + optimizer state → multi-A100 cluster 필요.
• Storage: 매 task 마다 full checkpoint 저장 시 비용 폭발.
• Catastrophic forgetting: 매 full FT 가 base capability 손상.
• PEFT: 매 base frozen, delta 만 학습 → 1 base + N tiny adapters.

매 family

• LoRA (Hu et al. 2021): low-rank decomposition ΔW = BA, rank r=4-64.
• QLoRA (Dettmers et al. 2023): 4-bit NF4 quantized base + LoRA adapters.
• Prefix Tuning (Li & Liang 2021): learnable prefix tokens prepended to keys/values.
• Prompt Tuning (Lester et al. 2021): learnable soft prompts at input.
• IA³ (Liu et al. 2022): scale activations via learned vectors (multiply, not add).
• Adapters (Houlsby et al. 2019): small bottleneck MLPs inserted between layers.
• DoRA (2024): magnitude + direction decomposition, LoRA 보다 우수.

매 LoRA 수학

• W' = W + αBA/r where B ∈ R^{d×r}, A ∈ R^{r×k}, r ≪ min(d,k).
• Trainable: 2dr params instead of dk. 매 d=k=4096, r=8 → 65k vs 16M (250× 감소).
• Inference: 매 merge W ← W + αBA/r → zero overhead.

매 응용

1. Domain adaptation (legal, medical LLM).
2. Instruction tuning (Alpaca-style).
3. Style transfer (FLUX LoRA for art style).
4. Multi-tenant serving (1 base + N customer LoRAs).

💻 패턴

LoRA with peft library

from peft import LoraConfig, get_peft_model, TaskType
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")
config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16, lora_alpha=32, lora_dropout=0.05,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
)
model = get_peft_model(base, config)
model.print_trainable_parameters()  # ~0.5% trainable

QLoRA (4-bit base + LoRA)

from transformers import BitsAndBytesConfig
import torch

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)
base = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-70B", quantization_config=bnb, device_map="auto",
)
model = get_peft_model(base, lora_config)  # 70B on single 48GB GPU

Save / load adapter only

model.save_pretrained("./my-lora")  # ~50MB, not 140GB

from peft import PeftModel
loaded = PeftModel.from_pretrained(base, "./my-lora")

Merge for inference

merged = model.merge_and_unload()  # W ← W + αBA/r
merged.save_pretrained("./merged-model")  # standard HF model, no peft dep

Multi-LoRA serving (vLLM)

# vLLM 0.6+ supports dynamic LoRA loading
from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

llm = LLM(model="meta-llama/Llama-3.1-8B", enable_lora=True, max_loras=8)
out = llm.generate(prompts, sampling_params,
    lora_request=LoRARequest("customer-42", 1, "./customer-42-lora"))

DoRA (2024)

config = LoraConfig(r=16, lora_alpha=32, use_dora=True,  # peft >= 0.10
    target_modules=["q_proj", "v_proj"])

Prompt tuning

from peft import PromptTuningConfig, PromptTuningInit
config = PromptTuningConfig(
    task_type=TaskType.CAUSAL_LM,
    prompt_tuning_init=PromptTuningInit.TEXT,
    num_virtual_tokens=20,
    prompt_tuning_init_text="Classify sentiment:",
    tokenizer_name_or_path="meta-llama/Llama-3.1-8B",
)

매 결정 기준

상황	Approach
1 GPU, large base (70B)	QLoRA
Multi-task, single base	LoRA + multi-adapter serving
Tiny VRAM, frozen base OK	Prompt tuning
Best quality, less compute saving	DoRA
Diffusion model style	LoRA (rank 4-32)
Production accuracy critical	Full FT (if 가능)

기본값: QLoRA (4-bit NF4 + r=16 LoRA on q/k/v/o projections).

🔗 Graph

• 부모: Fine-Tuning
• 변형: LoRA · QLoRA · DoRA
• 응용: Fine-tuning · Domain-Adaptation
• Adjacent: LLM_Optimization_and_Deployment_Strategies · RLHF

🤖 LLM 활용

언제: 매 single GPU 에서 large model fine-tune, multi-tenant LoRA serving, rapid task iteration. 언제 X: 매 base model 의 fundamental capability 변경 필요 (continued pretraining → full FT or full pretraining).

❌ 안티패턴

• Rank too low: r=1-2 → underfitting. 매 r=8-32 starting point.
• Wrong target modules: only q_proj/v_proj skip → degraded. 매 all attention + MLP modules 가 best.
• Forgetting alpha: 매 alpha=2r convention 무시 → unstable training.
• Saving full model: model.save_pretrained() on PeftModel 만 saves adapter. Don't merge unnecessarily.
• QLoRA + bf16 base: 매 NF4 quantization 의 redundant. 매 fp16 or bf16 base 둘 중 하나.

🧪 검증 / 중복

• Verified (HuggingFace peft docs, Hu et al. 2021 LoRA, Dettmers et al. 2023 QLoRA).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — PEFT family, LoRA/QLoRA patterns, decision matrix