LoftQ's picture
Update README.md
d63d2e9
---
license: mit
language:
- en
pipeline_tag: text-generation
tags:
- 'quantization '
- lora
---
# LoftQ Initialization
| [Paper](https://arxiv.org/abs/2310.08659) | [Code](https://github.com/yxli2123/LoftQ) | [PEFT Example](https://github.com/huggingface/peft/tree/main/examples/loftq_finetuning) |
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, `Llama-2-13b-hf-4bit-64rank`, is obtained from [LLAMA-2-13b](https://huggingface.co/meta-llama/Llama-2-13b-hf).
The backbone is under `LoftQ/Llama-2-13b-hf-4bit-64rank` and LoRA adapters are under the `subfolder='loftq_init'`.
## Model Info
### Backbone
- Stored format: `torch.bfloat16`
- Size: ~ 26 GiB
- Loaded format: bitsandbytes nf4
- Size loaded on GPU: ~6.5 GiB
### LoRA adapters
- rank: 64
- lora_alpha: 64
- 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.
```python
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
MODEL_ID = "LoftQ/Llama-2-13b-hf-4bit-64rank"
base_model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
torch_dtype=torch.bfloat16, # you may change it with different models
quantization_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16, # bfloat16 is recommended
bnb_4bit_use_double_quant=False,
bnb_4bit_quant_type='nf4',
),
)
peft_model = PeftModel.from_pretrained(
base_model,
MODEL_ID,
subfolder="loftq_init",
is_trainable=True,
)
# Do training with peft_model ...
```
## Experiment Results
We have conducted experiments on supervised fine-tuning of [GSM8K](https://huggingface.co/datasets/gsm8k)
and [WikiText-2](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-raw-v1).
| Model | Bits | Rank | LoRA Initial | GSM8K | WikiText-2 |
| -------------- | ---- | ---- | -------------------- | ----- | ---------- |
| LLAMA-2-13b | 16 | 64 | Gaussian + 0 | 45.3 | 5.12 |
| LLAMA-2-13b | 4 | 64 | Gaussian + 0 (QLoRA) | 39.9 | 5.22 |
| **LLAMA-2-13b** | 4 | 64 | LoftQ | 45.0 | 5.16 |
**Inference** Here is an example code for inference after the model has been fine-tuned on [GSM8K](https://huggingface.co/datasets/gsm8k).
```python
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
MODEL_ID = "LoftQ/Llama-2-13b-hf-4bit-64rank"
base_model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
torch_dtype=torch.bfloat16, # you may change it with different models
quantization_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16, # bfloat16 is recommended
bnb_4bit_use_double_quant=False,
bnb_4bit_quant_type='nf4',
),
)
peft_model = PeftModel.from_pretrained(
base_model,
MODEL_ID,
subfolder="gsm8k",
is_trainable=True,
)
# Do inference with peft_model ...
```
See the full code at our [Github Repo]((https://github.com/yxli2123/LoftQ))
## Citation
```bibtex
@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}
}
```