a project that builds on zeus, an end-to-end pytorch plugin developed by umich researchers to jointly minimize energy usage and training time using, in part, a multi-armed bandit to select batch size and a just-in-time energy profiler selecting batch size and gpu power limits according to the requirements of the job. we extended the profiler to consider additional parameters such as learning rate and dropout rate. built using python (pytorch). 

​

for compsci 243 (minlan yu): advanced computer networks, a grad-level class about distributed ML models. project done with helen cho and charumathi badrinath

​

code here, writeup here