Training large, long-context models at near-peak throughput when GPU memory is scarce.
ShadowSpill: A system for training within a GPU memory budget. It automatically plans what stays on the GPU and what spills to slower memory. Its planner jointly handles offloading, micro-batch sizing and scheduling, recomputation, and parallelism/sharding.
Playground: Dataflow Simulator, a simplified version of the ShadowSpill planner that uses theoretical hardware constants to show the benefits of effective offloading.
Hi, I'm Andrew Sheinberg. I'm a fourth-year Computer Science PhD student at Princeton University, advised by Professor Kai Li. Before Princeton, I studied Computer Science & Mathematics at Yale. I grew up in Barrington, RI.
I build systems software for machine learning. Accelerators keep getting faster, but their memory is expensive and the links to the rest of the system are slow by comparison, so the hard part is often deciding where data lives and when it moves. My recent work borrows classic operating-systems ideas, like working sets and prefetching, to train large, long-context models at near-peak throughput with a fraction of the GPU memory they would otherwise need.
Outside of research, I've linked some side projects I've worked on over the years. I also enjoy traveling, and I've linked a few of my favorite photos above.
Teaching Assistant (Princeton)
Feel free to email me at: asheinb@princeton.edu