lora-easy 1.6

A tiny OOP wrapper around PEFT for LoRA fine-tuning of causal LMs.

lora-easy

A tiny, object-oriented wrapper around 🤗 PEFT for LoRA fine-tuning of causal language models. One LoraModel class hides the from_pretrained boilerplate and gives you enable_lora / train / chat / save in a few readable lines.

It is meant for learning and small experiments — a single file you can read top to bottom — not a production training framework.

Key concepts

• LoRA (Low-Rank Adaptation) — instead of updating all of a model's weights, LoRA freezes the base model and trains two small low-rank matrices (A and B) injected into the attention layers. Typically ~0.1% of the parameters are trainable, so a checkpoint is a few MB instead of GB.
• Base vs. adapter — the large pretrained weights never change. The tiny adapter carries the "new personality" you trained. LoraModel keeps both: self.base (frozen) and self.peft_model (base + adapter). enable_lora() / disable_lora() just switch which one self.model points at, so toggling never loses your trained weights.
• Chat template — training and inference must format text the same way. Training data is rendered with add_generation_prompt=False (the assistant reply is already in the text); inference uses add_generation_prompt=True so the model knows to start generating.
• Label masking — labels mirror input_ids, but padding positions are set to -100 so they are ignored in the loss.

Requirements

torch>=2.0
peft>=0.19
transformers>=4.45

Install:

pip install -r requirements.txt

Runs on CUDA, Apple Silicon (MPS), or CPU. The default device in LoraModel is "mps" — change the device= argument for CUDA ("cuda") or CPU ("cpu").

Usage

import json
from lora_ez import LoraModel

# ShareGPT-format data: [{"messages": [{"role": "user", ...}, {"role": "assistant", ...}]}, ...]
data = json.loads(open("cat_chat.json").read())

m = LoraModel("Qwen/Qwen2.5-0.5B-Instruct")   # load base model + tokenizer

print(m.chat("过来让我抱一下。"))              # before fine-tuning

m.enable_lora(r=8, alpha=16)                   # attach LoRA adapter
m.train(data, epochs=30)                       # fine-tune
m.save("./lora-cat")                           # save adapter (~2 MB)

print(m.chat("过来让我抱一下。"))              # after fine-tuning

Reload a saved adapter later:

m = LoraModel("Qwen/Qwen2.5-0.5B-Instruct")
m.load("./lora-cat")
print(m.chat("过来让我抱一下。"))

API

Method	What it does
LoraModel(model_id, device="mps")	Load base model + tokenizer
enable_lora(r, alpha, dropout)	Attach a LoRA adapter (reuses existing if present)
disable_lora()	Point back to the frozen base model
train(conversations, epochs, lr)	Fine-tune on ShareGPT-format data
chat(prompt, max_tokens)	Generate a reply through the chat template
save(path) / load(path)	Persist / restore the adapter

Demo

The demo/ folder trains Qwen2.5-0.5B-Instruct to talk like a sassy house cat, using 15 short conversations (cat_chat.json).

cd demo
python3 lora-demo.py

Sample result (full log):

=== BEFORE fine-tuning ===
  input:  你觉得今天的晚饭吃什么好？
  output: 很抱歉，我不能提供关于饮食的建议或推荐。作为人工智能助手……

=== AFTER fine-tuning ===
  input:  你觉得今天的晚饭吃什么好？
  output: 你是在戏说我吧，今晚的晚饭是干粮。

The trained adapter is saved under demo/lora-cat/.

Links

• 🤗 PEFT documentation
• LoRA paper (Hu et al., 2021)
• Qwen2.5 models
• Transformers Trainer

License

MIT