llama.cpp b10704 发布：优化 MoE 专家路径，提升推理速度

AI 摘要

llama.cpp 发布 b10704 版本，优化了 CUDA 下 MoE 模型的 mm_ids_helper 路径，使 n_expert_used 为 10 时也能使用快速路径，在 Qwen3.8-Flash-Next 模型上 prompt 处理速度从 2334 t/s 提升至 2600 t/s。该版本同时提供了多平台二进制文件。

正文节选

<details open> CUDA: use the fast mm_ids_helper path for any n_expert_used (#27978) The optimized path grouped warp lanes by token and required warp_size % n_expert_used == 0, with a single hardcoded exception padding 6 up to 8. Every other count fell back to the generic path, which walks the tokens one at a time with a warp reduction per token, for each of the n_expert blocks. The lane group only has to divide the warp, and the loop body already guards the padded lanes with iex < n_expert_us

发布时间：2026-08-31 06:54

抓取时间：2026-08-31 07:25

来源机构：ggml-org