Awesome

Efficient Inference for Big Models

What's New

• 2022/07/31 (BMInf 2.0.0) BMInf can now be applied to any transformer-based model.
• 2021/12/21 (BMInf 1.0.0) Now the package no more depends on cupy and supports PyTorch backpropagation.
• 2021/10/18 We updated the generate interface and added a new CPM 2.1 demo.
• 2021/09/24 We publicly released BMInf on the 2021 Zhongguancun Forum (AI and Multidisciplinary Synergy Innovation Forum).

Note: README for BMInf-1 can be found in old_docs directory. Examples of CPM-1/2 and EVA will be published soon.

Overview

BMInf (Big Model Inference) is a low-resource inference package for large-scale pretrained language models (PLMs).

BMInf supports running models with more than 10 billion parameters on a single NVIDIA GTX 1060 GPU in its minimum requirements. Running with better GPUs leads to better performance. In cases where the GPU memory supports the large model inference (such as V100 or A100), BMInf still has a significant performance improvement over the existing PyTorch implementation.

If you use the code, please cite the following paper:

@inproceedings{han2022bminf,
	title={BMInf: An Efficient Toolkit for Big Model Inference and Tuning},
	author={Han, Xu and Zeng, Guoyang and Zhao, Weilin and Liu, Zhiyuan and Zhang, Zhengyan and Zhou, Jie and Zhang, Jun and Chao, Jia and Sun, Maosong},
	booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations},
	pages={224--230},
	year={2022}
}

Installation

• From pip: pip install bminf

• From source code: download the package and run python setup.py install

Hardware Requirement

Here we list the minimum and recommended configurations for running BMInf.

GPUs with compute capability 6.1 or higher are supported by BMInf. Refer to the table to check whether your GPU is supported.

Software Requirement

BMInf requires CUDA version >= 10.1 and all the dependencies can be automaticlly installed by the installation process.

• python >= 3.6
• torch >= 1.7.1
• cpm_kernels >= 1.0.9

Quick Start

Use bminf.wrapper to automatically convert your model.

import bminf

# initialize your model on CPU
model = MyModel()

# load state_dict before using wrapper
model.load_state_dict(model_checkpoint)

# apply wrapper
with torch.cuda.device(CUDA_DEVICE_INDEX):
    model = bminf.wrapper(model)

If bminf.wrapper does not fit your model well, you can use the following method to replace it manually.

• Replace torch.nn.ModuleList with bminf.TransformerBlockList.

module_list = bminf.TransformerBlockList([
	# ...
], [CUDA_DEVICE_INDEX])

• Replace torch.nn.Linear with bminf.QuantizedLinear.

linear = bminf.QuantizedLinear(torch.nn.Linear(...))

Performances

Here we report the speeds of CPM2 encoder and decoder we have tested on different platforms. You can also run benchmark/cpm2/encoder.py and benchmark/cpm2/decoder.py to test the speed on your machine!

Community

We welcome everyone to contribute codes following our contributing guidelines.

You can also find us on other platforms:

• QQ Group: 735930538
• WeChat Official Account: OpenBMB
• Website: https://www.openbmb.org
• Weibo: http://weibo.cn/OpenBMB
• Twitter: https://twitter.com/OpenBMB

License

The package is released under the Apache 2.0 License.

References

1. CPM-2: Large-scale Cost-efficient Pre-trained Language Models. Zhengyan Zhang, Yuxian Gu, Xu Han, Shengqi Chen, Chaojun Xiao, Zhenbo Sun, Yuan Yao, Fanchao Qi, Jian Guan, Pei Ke, Yanzheng Cai, Guoyang Zeng, Zhixing Tan, Zhiyuan Liu, Minlie Huang, Wentao Han, Yang Liu, Xiaoyan Zhu, Maosong Sun.
2. CPM: A Large-scale Generative Chinese Pre-trained Language Model. Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
3. EVA: An Open-Domain Chinese Dialogue System with Large-Scale Generative Pre-Training. Hao Zhou, Pei Ke, Zheng Zhang, Yuxian Gu, Yinhe Zheng, Chujie Zheng, Yida Wang, Chen Henry Wu, Hao Sun, Xiaocong Yang, Bosi Wen, Xiaoyan Zhu, Minlie Huang, Jie Tang.
4. Language Models are Unsupervised Multitask Learners. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever.