LoftQ/Llama-2-7b-hf-fp16-64rank-gsm8k

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/Llama-2-7b-hf-fp16-64rank-gsm8k, is LoRA fine-tuned from LLAMA-2-7b on GSM8K dataset.

Model Info

LoRA adapters

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

Usage

Inference Here is an example code for inference after the model has been fine-tuned on GSM8K.

import torch
from transformers import AutoModelForCausalLM

MODEL_ID = "LoftQ/Llama-2-7b-hf-fp16-64rank-gsm8k"

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    torch_dtype=torch.bfloat16,  # you may change it with different models
    token=YOUR_HF_TOKEN,
)
# you can also merge the LoRA adapters to the backbone if you like
model = model.merge_and_unload()

# Do inference with `model` ...

See full evaluation on GSM8K on Github.

Experiment Results

We have conducted experiments on supervised fine-tuning of GSM8K and WikiText-2.

Model	Bits	Rank	LoRA Initial	GSM8K
LLAMA-2-7b	16	64	Gaussian + 0	36.9
LLAMA-2-7b	4	64	Gaussian + 0 (QLoRA)	35.1
LLAMA-2-7b	4	64	LoftQ	35.0

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}
}