RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed).

OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models

🔥🔥🔥

OmniQuant is a simple and powerful quantization technique for LLMs. The current release supports:

OmniQuant algorithm for accurate weight-only quantization (W4A16/W3A16/W2A16) and weight-activation quantization (W6A6, W4A4)
Pre-trained Omniquant model zoo for LLMs (LLaMA-1&2, LLaMA-2-Chat, OPT, Falcon, Mixtral-7Bx8; load to generate quantized weights).
A out-of-the-box case that leverages MLC-LLM to run LLaMa-2-Chat (7B/13B) with W3A16g128 quantization on GPUs and mobile phones.

News

[2024/7] 🔥 We release a new quantization algorithm, EfficientQAT, which realizes quantization-aware training in a time-efficient and memory-efficient manner. Additionally, EfficientQAT is the current SoTA of uniform quantization.
[2024/1] 🌟 Our OmniQuant paper has been accepted for a Spotlight presentation at ICLR 2024 (only top 5% out of over 7200 submissions)! 🎉 Cheers!
[2023/12] 🔥 We provide support for Mixtral-8x7B. OmniQuant is capable of achieving near-lossless 4-bit quantization with Mixtral-8x7B-v0.1, which reduces the memory requirement from 87GB to 23GB. Notably, the work-in-progress OmniQuant v2 is anticipated to outperform OmniQuant v1. We invite you to stay tuned for the upcoming, more powerful OmniQuant v2. You can access the model URL and the inference code for the quantized models at runing_quantized_mixtral_7bx8.
[2023/09] 🔥 We have expanded support for Falcon. OmniQuant efficiently compresses Falcon-180b from 335G to 65G, with minimal performance loss. Furthermore, this compression allows for Falcon-180b inference on a single A100 80GB GPU. For details, refer to runing_falcon180b_on_single_a100_80g.

Install

conda create -n omniquant python=3.10 -y
conda activate omniquant
git clone https://github.com/OpenGVLab/OmniQuant.git
cd OmniQuant
pip install --upgrade pip 
pip install -e .

We also leverage the kernel from AutoGPTQ to achieve real quantization. So you should also install the bug-fixed AutoGPTQ as follows::

git clone https://github.com/ChenMnZ/AutoGPTQ-bugfix
pip install -v .

OmniQuant Model Zoo

We provide pre-trained Omniquant model zoo for multiple model families, including LLaMa-1&2, LLaMa-2-Chat, OPT.

You can download the pre-trained OmniQuant parameters you need at Huggingface.

The detailed support list:

Models	Sizes	W2A16	W2A16g128	W2A16g64	W3A16
LLaMA	7B/13B/30B/65B	✅	✅	✅	✅
LLaMA-2	7B/13B/70B	✅	✅	✅	✅
OPT	125m/1.3B/2.7B/6.7B/13B/30B/66B	✅	✅	✅	✅

Models	Sizes	W3A16g128	W4A16	W4A16g128	W6A6	W4A4
LLaMA	7B/13B/30B/65B	✅	✅	✅	✅	✅
LLaMA-2	7B/13B/70B	✅	✅	✅	✅	✅
OPT	125m/1.3B/2.7B/6.7B/13B/30B/66B	✅	✅	✅	✅	✅
LLaMA-2-Chat	7B/13B	✅

Usage

We provide full script to run OmniQuant in ./scripts/. We use LLaMa-7B as an example here:

Obtain the channel-wise scales and shifts required for initialization:

conda install git git-lfs
git lfs install
git clone https://huggingface.co/ChenMnZ/act_shifts
git clone https://huggingface.co/ChenMnZ/act_scales

Optional, we also offer the script that you can generate channel-wise scales and shifts by yourself:

python generate_act_scale_shift.py --model /PATH/TO/LLaMA/llama-7b

Weight-only quantization

# W3A16
CUDA_VISIBLE_DEVICES=0 python main.py \
--model /PATH/TO/LLaMA/llama-7b  \
--epochs 20 --output_dir ./log/llama-7b-w3a16 \
--eval_ppl --wbits 3 --abits 16 --lwc

# W3A16g128
CUDA_VISIBLE_DEVICES=0 python main.py \
--model /PATH/TO/LLaMA/llama-7b  \
--epochs 20 --output_dir ./log/llama-7b-w3a16g128 \
--eval_ppl --wbits 3 --abits 16 --group_size 128 --lwc

weight-activation quantization

# W4A4
CUDA_VISIBLE_DEVICES=0 python main.py \
--model /PATH/TO/LLaMA/llama-7b  \
--epochs 20 --output_dir ./log/llama-7b-w4a4 \
--eval_ppl --wbits 4 --abits 4 --lwc --let \
--tasks piqa,arc_easy,arc_challenge,boolq,hellaswag,winogrande

reproduce evaluation results of our paper

1) download the pretrained OmniQuant parameters you want through Huggingface.

2) set epoch as 0 and inference with resume, take LLaMa-7B with W3A16g128 quantization as an example:

CUDA_VISIBLE_DEVICES=0 python main.py \
--model /PATH/TO/LLaMA/llama-7b  \
--epochs 0 --output_dir ./log/test \
--eval_ppl --wbits 3 --abits 16 --group_size 128 --lwc \
--resume /PATH/TO/Pretrained/Parameters

More detailed and optional arguments:

--model: the local model path or huggingface format.
--wbits: weight quantization bits.
--abits: activation quantization bits.
--group_size: group size of weight quantization. If no set, use per-channel quantization for weight as default.
--lwc: activate the Learnable Weight Clipping (LWC).
--let: activate the Learnable Equivalent Transformation (LET).
--lwc_lr: learning rate of LWC parameters, 1e-2 as default.
--let_lr: learning rate of LET parameters, 5e-3 as default.
--epochs: training epochs. You can set it as 0 to evaluate pre-trained OmniQuant checkpoints.
--nsamples: number of calibration samples, 128 as default.
--eval_ppl: evaluating the perplexity of quantized models.
--tasks: evaluating zero-shot tasks.
--resume: loading pre-trained OmniQuant parameters.
--multigpu: to inference larger network on multiple GPUs
--real_quant: real quantization, which can see memory reduce. Note that due to the limitations of AutoGPTQ kernels, the real quantization of weight-only quantization can only lead memory reduction, but with slower inference speed.
--save_dir: saving the quantization model for further exploration.

Runing Quantized Models with MLC-LLM

MLC-LLM offers a universal deployment solution suitable for various language models across a wide range of hardware backends, encompassing iPhones, Android phones, and GPUs from NVIDIA, AMD, and Intel.

We compile the OmniQuant's quantization models through MLC-LLM and offer an out-of-the-box case here. You can see smaller gpu memory usage and inference speedup. Detailed instructions can be found in in runing_quantized_models_with_mlc_llm.ipynb.

Specially, we also deploy the aforementioned two quantized models into mobile phones through MLC-LLM. You can download the Android app by simply clicking the button below:

This app includes three models, LLaMa-2-7B-Chat-Omniquant-W3A16g128asym, LLaMa-2-13B-Chat-Omniquant-W3A16g128asym, and LLaMa-2-13B-Chat-Omniquant-W2A16g128asym. They require at least 4.5G, 7.5G, and 6.0G free RAM, respectively. Note that 2bit quantization has worse performance compared to 3bit quantization as shown in our paper. The inclusion of 2-bit quantization is just an extreme exploration about deploy LLM in mobile phones. Currently, this app is in its demo phase and may experience slower response times, so wait patiently for the generation of response. We have tested this app on Redmi Note 12 Turbo (Snapdragon 7+ Gen 2 and 16G RAM), some examples are provided below:

LLaMa-2-7B-Chat-Omniquant-W3A16g128asym

LLaMa-2-13B-Chat-Omniquant-W3A16g128asym

LLaMa-2-13B-Chat-Omniquant-W2A16g128asym

We also have tested this app on iPhone 14 Pro (A16 Bionic and 6G RAM), some examples are provided below:

LLaMa-2-7B-Chat-Omniquant-W3A16g128asym

Results

OmniQuant achieve SoTA performance in weight-only quantization
OmniQuant achieve SoTA performance in weight-activation quantization
OmniQuant is generalize, also obatins excellent performance in instruction-tuned models with GPT-4 evaluation
MLC-LLM can obtain really speedup and memory saving for W4A16/W3A16/W2A16 quantization

Related Project

SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models

AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration

GPTQ: Accurate Post-training Compression for Generative Pretrained Transformers

RPTQ: Reorder-Based Post-Training Quantization for Large Language Models

MLC LLM

AutoGPTQ

Citation

If you use our OmniQuant approach in your research, please cite our paper:

@article{OmniQuant,
  title={OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models},
  author={Shao, Wenqi and Chen,Mengzhao and  Zhang, Zhaoyang and Xu, Peng and Zhao, Lirui and Li, Zhiqian and Zhang, Kaipeng Zhang, and Gao, Peng, and Qiao, Yu, and Luo, Ping},
  journal={arXiv preprint arXiv:2308.13137},
  year={2023}
}

	if args.aug_loss:
	loss += loss_func(fp_inps_2[index:index+args.batch_size,], quant_out)

opengvlab / omniquant Goto Github PK

omniquant's Introduction

OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models

🔥🔥🔥

News

Contents

Install

OmniQuant Model Zoo

Usage

Runing Quantized Models with MLC-LLM

Results

Related Project

Citation

omniquant's People

Contributors

Stargazers

Watchers

Forkers

omniquant's Issues

SOLAR-10.7B, is a compact, yet remarkably powerful large language model; it has demonstrated unparalleled state-of-the-art performance in models under 30B parameters - rivaling model's with up to 30B parameters in performance.

This is -- a caller that pip uses to run setup.py

- It imports setuptools before invoking setup.py, to enable projects that directly

import from distutils.core to work with newer packaging standards.

- It provides a clear error message when setuptools is not installed.

- It sets sys.argv[0] to the underlying setup.py, when invoking setup.py so

setuptools doesn'"'"'t think the script is -c. This avoids the following warning:

manifest_maker: standard file '"'"'-c'"'"' not found".

- It generates a shim setup.py, for handling setup.cfg-only projects.

Recommend Projects

Recommend Topics

Recommend Org

import from `distutils.core` to work with newer packaging standards.

- It sets `sys.argv[0]` to the underlying `setup.py`, when invoking `setup.py` so

setuptools doesn'"'"'t think the script is `-c`. This avoids the following warning: