[Perf] ~2x decode throughput regression for structured outputs since #45424: apply_grammar_bitmask staging rewrite (bisected to commit, file
![[Perf] ~2x decode throughput regression for structured outputs since #45424: apply_grammar_bitmask staging rewrite (bisected to commit, file](https://www.chat-gpts.plus/wp-content/uploads/2026/09/49013-aa38f76f-768x403.jpg)
该问题通常出现在 vLLM v0.24.0 / v0.25.1 上运行 xgrammar 结构化输出(guided JSON)时,引擎每步在 V1/V2 runner 里重新分配并 pin 一份 12.6 MB 级别的 bitmask staging buffer,导致含结构化输出的 batch 解
![[Bug]: Reasoning still returned in /responses while include_reasoning is set to false](https://www.chat-gpts.plus/wp-content/uploads/2026/09/56428-09f41fe1-768x403.jpg)
![[Performance]: MTP first repeat misses prefix cache on a hybrid Mamba/GDN model](https://www.chat-gpts.plus/wp-content/uploads/2026/09/53504-fd99d934-768x403.jpg)






