Reproduce “Elasticy: A Runtime-Reconfigurable Architecture for Dynamic Streams” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Konstantin J Hossfeld, Nathan Sobotka, Gina Sohn, Tun-Yu Chang, Kunle Olukotun
Elasticy targets runtime-varying tensor shapes and control flow in accelerator workloads, including LLM inference. It processes tiles in streams, using rank-variable stop tokens to mark tensor boundaries and distributed schedulers to reconfigure at stream boundaries. For ragged Llama 3-8B decode, it reaches 92.59% roofline attainment versus 69.67% for H100; Elasticy results come from simulation, not fabricated hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.