Paper | Usage | Citation | Video |
This repository contains the official code for our ICLR 2023 spotlight paper MPCFormer: fast, performant, and private transformer inference with MPC. We design MPCFormer to protect users' data privacy by using Secure Multiparty Computation(MPC). It also meets other real-world requirements:
- inference latency: by replacing bottleneck functions by their MPC-friendly ones.
- ML performance: by introducing a subsequent Knowledge-Distillation(KD) procedure.
It achieves 5.26x speedup for Bert-Base MPC inference, while preserving a similar ML accuracy. More comprehensive results such as on Bert-Large, Roberta, can be found in the paper.
To install necessary packages, install the transformer directory in editor mode:
git clone https://github.com/MccRee177/MPCFormer
cd MPCFormer/transformers
pip install -e .
Step 1: Obtain a teacher Transformer model by fine-tuning on downstream tasks Here
We support GLUE and Imdb, other datasets can be easily supported via the ransformers library.
Step 2: perform approximation and distillation of MPCFormer Here.
(Optional) evaluate baselines in the paper Here.
(Optional) Benchmark the inference time of approximated model: Here.
If you find this repository useful, please cite our paper using
@article{li2022mpcformer,
title={MPCFormer: fast, performant and private Transformer inference with MPC},
author={Li, Dacheng and Shao, Rulin and Wang, Hongyi and Guo, Han and Xing, Eric P and Zhang, Hao},
journal={arXiv preprint arXiv:2211.01452},
year={2022}
}