AI 中文总结
研究针对扩散大语言模型服务的收敛异质性问题，提出BlockServe框架，集成块粒度调度、混合状态执行及计算感知准入控制器，在五个基准测试中实现高吞吐量，确立块粒度调度为高吞吐量离线dLLM推理基础。
AI 中文摘要
扩散大语言模型（dLLMs）的高效服务受到收敛异质性的阻碍：批处理多个请求时，不同序列以不同速率收敛，导致较快请求在较慢请求后停滞，产生计算气泡和尾部延迟。我们提出了BlockServe，这是一个连续批处理框架，它将块粒度调度（在块边界立即淘汰已完成请求）与混合状态执行相结合，通过收集-散射索引将双缓存和并行解码扩展到异构批处理。此外，一个计算感知准入控制器通过令牌预算补充来扩展有效批处理容量。在五个基准测试中的Dream和LLaDA上，BlockServe在生成质量可比的情况下，吞吐量比Fast-dLLM高1.9至10.6倍，确立了块粒度调度作为高吞吐量离线dLLM推理的基础。
英文摘要
Efficient serving of diffusion large language models (dLLMs) is hindered by convergence heterogeneity: when batching multiple requests, different sequences converge at different rates, causing faster requests to stall behind slower stragglers and introducing compute bubbles and tail latency. We present BlockServe, a continuous batching framework that integrates block-grained scheduling -- immediately evicting completed requests at block boundaries -- with mixed-state execution that extends dual cache and parallel decoding to heterogeneous batches via gather-scatter indexing. Furthermore, a compute-aware admission controller expands effective batch capacity through token-budgeted refill. On Dream and LLaDA across five benchmarks, BlockServe achieves 1.9--10.6$\times$ throughput over Fast-dLLM with comparable generation quality, establishing block-grained scheduling as a foundation for high-throughput offline dLLM inference.
