HiSparse wheel 0.24.1.dev0+hisparse.9bb4bea.cu129
vLLM 0.24-based HiSparse wheel for GLM-5.x PD decode instances (sm90a, cp312, CUDA 12.9 toolchain / torch cu128-compatible). Built from feat/hisparse-mla-decode @ 9bb4bea — host-resident sparse-MLA decode hot-buffering with plan-once index sharing and overlapped prefetch (default on). See vllm-project#46326 for the full PR and measured results.