Logion participates in Princeton Open Hackathon

This June 11-13, Logion participated in Princeton’s Open Hackathon, an event jointly held by Princeton, Nvidia, and Open ACC Organization. The Princeton Open Hackathon is a multi-day event intended to assist computational research projects in accelerating and optimizing their code, specifically to run on Nvidia GPUs.

The Logion team at the Hackathon consisted of: Jacob Murel (Research Software Engineer for the Logion Project), Sepideh Khajehei (Nvidia), Mohamed Abdellatif (Princeton Research Computing), Anushka Acharya (Princeton Research Computing), and Mattie Niznik (Princeton Research Computing). Together, they optimized code for Logion’s error detection and correction algorithm.

Their efforts led to an approximately 6x speed-up in Logion’s error detection code on the Nvidia A100 and H100 GPUs used on Princeton’s HPC cluster. According to the Hackathon’s evaluative metrics, this equates to saving approximately 217,000 kWh per year (~$39,000 per year) if Logion’s error detection code ran 24/7/365 on Princeton’s cluster.

While the Hackathon team largely focused on runtime optimization, they nevertheless refactored code to improve the quality of Logion’s results. Specifically, the team implemented a blacklist of words and characters through which model results are filtered. These words and characters had been previously identified by Logion-affiliated philologists—notably Johannes Haubold—as persistently inaccurate and unsound conjectures. Improvements like this show how the Logion Project uses philological insights to inform language model code designed for premodern Greek.

Both the code speed-up and fine-tuned results will improve user experience of the Logion app. Currently, the Logion desktop application is undergoing user testing. Its web version, which is designed to run on Princeton’s GPU cluster, is in development.