BlockLock: Trade Fixed Geometric Constraints for Output-Level Reproducibility under Shared Batch Prefilling

arXiv:2611.06977 (submitted on 2026-11-04) proposes BlockLock, which preserves the geometry observed by protected requests during shared-batch prefilling, introducing two modes: BlockLock-Rows and BlockLock-Pack. Initial tests on five dense models with single and dual-card tensor parallelism using A6000, A100, and H100 report that the target sequence matches the bits of the top-20 logits. Under equivalent delivered workloads, BlockLock-Rows reduces latency by approximately 53.5%–55.0% and increases throughput by approximately 115.2%–122.2% compared to vLLM Batch-Invariant, without relying on whole-batch custom batch-invariant kernels.