Reproduce “BlockLock: Reproducible LLM Inference with Fixed Batch Geometry” 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
Xiuyuan Qi, Pedro Reviriego, Javier Conde, Shanshan Liu
BlockLock makes a target request’s inference repeatable across companion-perturbed batches by fixing its execution geometry within a deployment. Its Rows and Pack modes retain shared prefill while keeping the target’s layout fixed, with different decode work. Across 1,316 single-GPU conditions, both achieved 100% sequence and Top-20 logit agreement, unlike padded batching; this does not establish determinism across deployments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.