metadata

license: mit
language:
  - en
pipeline_tag: text-generation
tags:
  - 'quantization '
  - lora
  - loftq
  - llama

LoftQ Initialization

| Paper | Code | PEFT Example |

LoftQ (LoRA-fine-tuning-aware Quantization) provides a quantized backbone Q and LoRA adapters A and B, given a full-precision pre-trained weight W.

This model, LoftQ/Meta-Llama-3-8B-Instruct-4bit-64rank, is obtained from Llama-3-8B-Instruct. The backbone is under LoftQ/Meta-Llama-3-8B-Instruct-4bit-64rank and LoRA adapters are under the subfolder='loftq_init'.

Model Info

Backbone

Size: ~ 6 GiB
Loaded format: bitsandbytes nf4
Size loaded on GPU: ~ 6 GiB

LoRA adapters

rank: 64
lora_alpha: 16
target_modules: ["down_proj", "up_proj", "q_proj", "k_proj", "v_proj", "o_proj", "gate_proj"]

Usage

Training

Here's an example of loading this model and preparing for the LoRA fine-tuning.

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

MODEL_ID = "LoftQ/Meta-Llama-3-8B-Instruct-4bit-64rank"

base_model = AutoModelForCausalLM.from_pretrained(MODEL_ID)
peft_model = PeftModel.from_pretrained(
    base_model,
    MODEL_ID,
    subfolder="loftq_init",
    is_trainable=True,
)

# Do training with peft_model ...

See the full code at our Github Repo

Citation

@article{li2023loftq,
  title={Loftq: Lora-fine-tuning-aware quantization for large language models},
  author={Li, Yixiao and Yu, Yifan and Liang, Chen and He, Pengcheng and Karampatziakis, Nikos and Chen, Weizhu and Zhao, Tuo},
  journal={arXiv preprint arXiv:2310.08659},
  year={2023}
}