Reproduce “Local Inference of LLMs on Long Contexts: A Data Systems Perspective” 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
Mattias Dürrmeier, Alberto Lerner
The authors profile single-user decoding in a repeated decoder block across two model families, three NVIDIA GPU generations, and contexts up to one million tokens. For Qwen3-8B, they report attention uses 21% of Blackwell’s bandwidth at long contexts, while Turing and Ada approach saturation. These are kernel measurements, not a full serving benchmark or demonstrated optimization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.