This repo supports the paper "QLoRA: Efficient Finetuning of Quantized LLMs", an effort to democratize access to LLM research.
| Paper | Adapter Weights | Demo | Getting Started |
QLoRA uses bitsandbytes for quantization and is integrated with Huggingface's PEFT and transformers libraries. QLoRA was developed by members of the University of Washington's UW NLP group.
We present QLoRA, an efficient finetuning approach that reduces memory usage enough to finetune a 65B parameter model on a single 48GB GPU while preserving full 16-bit finetuning task performance. QLoRA backpropagates gradients through a frozen, 4-bit quantized pretrained language model into Low Rank Adapters (LoRA). Our best model family, which we name Guanaco, outperforms all previous openly released models on the Vicuna benchmark, reaching 99.3% of the performance level of ChatGPT while only requiring 24 hours of finetuning on a single GPU. QLoRA introduces a number of innovations to save memory without sacrificing performance: (a) 4-bit NormalFloat (NF4), a new data type that is information theoretically optimal for normally distributed weights (b) Double Quantization to reduce the average memory footprint by quantizing the quantization constants, and (c) Paged Optimizers to manage memory spikes. We use QLoRA to finetune more than 1,000 models, providing a detailed analysis of instruction following and chatbot performance across 8 instruction datasets, multiple model types (LLaMA, T5), and model scales that would be infeasible to run with regular finetuning (e.g. 33B and 65B parameter models). Our results show that QLoRA finetuning on a small high-quality dataset leads to state-of-the-art results, even when using smaller models than the previous SoTA. We provide a detailed analysis of chatbot performance based on both human and GPT-4 evaluations showing that GPT-4 evaluations are a cheap and reasonable alternative to human evaluation. Furthermore, we find that current chatbot benchmarks are not trustworthy to accurately evaluate the performance levels of chatbots. We release all of our models and code, including CUDA kernels for 4-bit training.
We release the resources associated with QLoRA finetuning in this repository under MIT license. In addition, we release the Guanaco model family for base LLaMA model sizes of 7B, 13B, 33B, and 65B. These models are intended for purposes in line with the LLaMA license and require access to the LLaMA models.
Access the live demo at the following link (coming soon).
In the meantime, we also provide a Colab notebook for easy comparison of ChatGPT-3.5 and Guanaco 65B model reponses to Vicuna prompts. You can access the Colab here or find it in the eval/
folder. Setting the display_model_names=True
flag and runing on all cells displays the model names.
Due to resource constraints the demo could be slow. We are working to release fast inference kernels to alleviate inference speed issues.
Coming soon.
The necessary changes to use QLoRA will soon be merged in the HF transformers library. We will update with installation info as soon as possible.
The qlora.py
code is a starting point for finetuning and inference on various datasets.
Basic command for finetuning a baseline model on the Alpaca dataset:
python qlora.py --model_name_or_path <path_or_name>
For models larger than 13B, we recommend adjusting the learning rate:
python qlora.py –learning_rate 0.0001 --model_name_or_path <path_or_name>
Quantization parameters are controlled from the BitsandbytesConfig
(see HF documenation) as follows:
- Loading in 4 bits is activated through
load_in_4bit
- The datatype used for the linear layer computations with
bnb_4bit_compute_dtype
- Nested quantization is activated through
bnb_4bit_use_double_quant
- The datatype used for qunatization is specified with
bnb_4bit_quant_type
. Note that there are two supported quantization datatypesfp4
(four bit float) andnf4
(normal four bit float). The latter is theoretically optimal for normally distributed weights and we recommend usingnf4
.
model = AutoModelForCausalLM.from_pretrained(
model_name_or_path='/name/or/path/to/your/model',
load_in_4bit=True,
device_map='auto',
max_memory=max_memory,
torch_dtype=torch.bfloat16,
quantization_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type='nf4'
),
)
You can access the paged optimizer with the argument --optim paged_adamw_32bit
We provide generations for the models described in the paper for both OA and Vicuna queries in the eval/generations
folder. These are intended to foster further research on model evaluation and analysis.
We also provide a Colab notebook for easy comparison of ChatGPT-3.5 and Guanaco 65B model reponses to Vicuna prompts. You can access the Colab here or find it in the eval/
folder. Setting the display_model_names=True
flag and runing on all cells displays the model names.
We include scripts adapted from the FastChat repo to automatically evaluate model generations using GPT-4. We include script for comparisons relative to ChatGPT with scores out of 10 as well as "pairwise comparisons" with three class labeling (win, loose, or tie). These are found in the eval
folder.
To facilitate the replication of our evaluation and future work in this area, we release GPT-4 and human ratings of our systems. These are found under eval/ratings-human
and eval/ratings-gpt4
.
More details can be found at eval/EVAL_README.md
.
Here a list of known issues and bugs. If your issue is not reported here, please open a new issue and describe the problem.
- 4-bit inference is slow. Currently, our 4-bit inference implementation is not yet integrated with the 4-bit matrix multiplication
- Resuming a LoRA training run with the Trainer currently runs on an error
- Currently, using
bnb_4bit_compute_type='fp16'
can lead to instabilities. For 7B LLaMA, only 80% of finetuning runs complete without error. We have solutions, but they are not integrated yet into bitsandbytes.
@article{dettmers2023qlora,
title={QLoRA: Efficient Finetuning of Quantized LLMs},
author={Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke},
journal={arXiv preprint arXiv:2305.14314},
year={2023}
}
This repo builds on the Stanford Alpaca and LMSYS FastChat repos.