[Adapt] [seminar] Accelerating Inference for Large Pre-trained Language Models

[任思宇]

Hi Adapters,


As large-scale pre-trained language models are ubiquitously used in various NLP tasks, the tremendous computational cost incurred by such models hinders their practicality in resource-constrained and time-sensitive scenarios. Therefore, a plethora of acceleration methods emerged by tackling different aspects of model redundancy. In this talk, I will introduce two categories of model acceleration algorithms: sample-adaptive early-exit and layer-adaptive length reduction. The former is built upon the idea that different samples have different difficulty, while the latter focus on in-sample redundancy among tokens.


Time: 2022/04/06 16:00-18:00


Venus: 
Tencent Meeting Link: https://meeting.tencent.com/dm/qDcCYYV2YUIj
Tencent Meeting Number：202-706-320


Hope you will find it useful and interesting!


Best regards,
Roy
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs.sjtu.edu.cn/pipermail/adapt/attachments/20220406/edd1d570/attachment.htm>