返回全部动态
llama.cpp b10731 发布：支持 Qwen4Exp 循环状态回滚，提升 MTP 解码速度
原标题：b10731
AI 摘要
llama.cpp 发布 b10731 版本，新增对 Qwen4Exp 模型循环状态回滚的支持，以优化 MTP 投机解码性能。该更新通过为每个槽位写入快照，修复了卷积状态恢复问题，使解码速度在代码和散文场景下分别提升至 183 和 144 tok/s，相比之前有显著提升。
以上摘要由 AI 生成，可能存在误差。事实请以原文为准。
正文节选
<details open> qwen4exp: support recurrent state rollback (#28123) MTP speculative decoding needs the target state to move back by the number of rejected draft tokens. Without rollback support the context is classified as SEQ_RM_TYPE_FULL and the server serializes the whole recurrent state to host memory on every round, which costs more than the drafting saves. The recurrent cache already holds n_rs_seq + 1 snapshot planes and the delta net writes its SSM state into them, but build_conv_state
发布时间：2026-09-01 13:34
抓取时间：2026-09-01 13:44
来源机构：ggml-org