llama.cpp b11265：Vulkan MoE 感知 tile 选择优化
原标题：b11265
AI 摘要
llama.cpp 发布 b11265 版本，主要修复 Vulkan 后端中 MoE 感知的 mat_mul_id tile 选择问题。此前 tile 选择依据总 token 数，而 MoE 分发网格中每个 workgroup 的真实 N 是每专家行数，导致在 Sarvam 30B 上 pp128 场景下实际仅约 6 行活跃却选了 l-tile，多数 worker 空转，拖慢的部分占整个任务 55%。该版本同时提供 macOS、Linux、Windows、Android 等多平台预编译包。
正文节选
<details open> vulkan: MOE aware mat_mul_id tile selection (#29182) mut_mul_id selected its matmul tile with total token count. For MoE dispatch grid the true N per workgroup is per-expert rows. At pp128 on Sarvam 30B that is 6, not 128, so the picker took the l-tile for ~6 live rows. Most workers in each group had nothing to do. This wasted time. The slow part was 55% of the whole job. </details> **Website:** - <https://llama.app> **Attestations:** - <https://github.com/ggml-org/llama.cpp/