BlockLock: Reproducible LLM Inference with Fixed Batch Geometrywith Xiuyuan Qi, Pedro Reviriego, Javier Conde1 versionW_afuhx6hfv1 · current