Anand Iyer

ReInc: Scaling Training of Dynamic Graph Neural Networks Dynamic Graph Neural Networks (DGNNs) have gained widespread attention due to their applicability in diverse domains such as traffic network prediction, epidemiological forecasting, and social network...

Dynamic Graph Neural Networks (DGNNs) can be really useful in many domains, but they are hard to scale. Our new work shows how to achieve an *order of magnitude* speedup when training DGNNs! Joint work with awesome folks Mingyu Guan and Taesoo Kim! arxiv.org/abs/2501.15348

OMEGA: A Low-Latency GNN Serving System for Large Graphs Graph Neural Networks (GNNs) have been widely adopted for their ability to compute expressive node representations in graph datasets. However, serving GNNs on large graphs is challenging due to the hi...

Serving GNNs on large graphs is challenging due to the high communication, computation, and memory overheads. Our new work, Omega, shows how to solve this. Joint work with awesome collaborators @daehyeok_kim, @adityaakella and amazing students! arxiv.org/abs/2501.08547

New Research Enables Faster, More Efficent Machine Learning Models | College of Computing

Our recent @TheOfficialACM SOSP papers on accelerating LLM inference are featured on @gtcomputing news: www.cc.gatech.edu/news/new-res...! Interested to work on similar systems? @gatech_scs PhD application deadline is tomorrow (Dec 15): scs.gatech.edu/phd-programs #AcademicSky

We are fast approaching the PhD application deadline this cycle (Dec 15 for @GeorgiaTech @gatech_scs, we're hiring so do consider us if you're interested in systems!), good luck to all applicants!! #academicsky #phdsky

It's PhD application season again. I'll be on the look out for motivated students. Located in the vibrant city of Atlanta, @GeorgiaTech @gatech_scs is a top 5 school with an awesome set of people. Apply by Dec 15 and mention my name! scs.gatech.edu/phd-programs

In India, it's commonly known as "ajinomoto", which I believe is a popular brand :)

MoEs promise reduced parameter activation, but in practice invoke all experts during batched inference. With my awesome students @vima_gupta, Kartik Sinha & collaborator Ada Gavrilovska, we show how MoEs can fulfill their promise in a workload-agnostic way: t.co/5nfGtxdOKb