#batched prefill

BlockLock: Reproducible LLM Inference with Fixed Batch GeometryNEW⭐9

BlockLock keeps LLM outputs bitwise reproducible under changing batches by fixing only the protected request's batch geometry, preserving shared prefill while c...