peft-fine-tuning

1# PEFT (Parameter-Efficient Fine-Tuning)
2 
3Fine-tune LLMs by training <1% of parameters using LoRA, QLoRA, and 25+ adapter methods.
4 
5## When to use PEFT
6 
7**Use PEFT/LoRA when:**
8- Fine-tuning 7B-70B models on consumer GPUs (RTX 4090, A100)
9- Need to train <1% parameters (6MB adapters vs 14GB full model)
10- Want fast iteration with multiple task specific adapters
11- Deploying multiple fine-tuned variants from one base model
12 
13**Use QLoRA (PEFT + quantization) when:**
14- Fine-tuning 70B models on single 24GB GPU
15- Memory is the primary constraint
16- Can accept ~5% quality trade-off vs full fine-tuning
17 
18**Use full fine-tuning instead when:**
19- Training small models (<1B parameters)
20- Need maximum quality and have compute budget
21- Significant domain shift requires updating all weights
22 
23## Quick start
24 
25### Installation
26 
27```bash
28# Basic installation
29pip install peft
30 
31# With quantization support (recommended)
32pip install peft bitsandbytes
33 
34# Full stack
35pip install peft transformers accelerate bitsandbytes datasets
36```
37 
38### LoRA fine-tuning (standard)
39 
40```python
41from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
42from peft import get_peft_model, LoraConfig, TaskType
43from datasets import load_dataset
44 
45# Load base model
46model_name = "meta-llama/Llama-3.1-8B"
47model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
48tokenizer = AutoTokenizer.from_pretrained(model_name)
49tokenizer.pad_token = tokenizer.eos_token
50…