[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1a56dngfam58x":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":12,"bodyZh":13,"bodyEn":14,"sourceName":15,"sourceUrl":16,"updatedAt":17},"llamacpp-b12059-unified-kv-multiseq-2026-11-03","RUNTIME","2026-11-02T23:04:48.717Z","llama.cpp b12059 改进多序列统一 KV：优化的是并发注意力，不是显卡通用倍速","llama.cpp b12059 improves unified KV for multiple sequences, not every GPU workload","新构建纳入 flash_attn_ext_rows，以紧凑的 KV 行索引减少多序列注意力的额外处理。上游旧测试不是本次新实测，不能推成个人电脑的提速承诺。","The new build includes compact KV row indexing for multi-sequence attention. Earlier upstream measurements are not fresh independent benchmarks or a general speedup promise.",[],[],"本次时间窗内，llama.cpp 发布 b12059，纳入多序列统一 KV 缓存的注意力改造。发布页时间为 2026 年 11 月 2 日 22:41 UTC，对应北京时间 11 月 3 日清晨。这是引擎代码的新发布，不是新模型，也不是 AraoAI 完成了一轮硬件复现。\n\n1. 它解决什么。多个请求共用统一 KV 缓存时，注意力计算需要排除不属于当前序列的 token。上游 PR 说明，这种过滤会带来额外开销；新增 flash_attn_ext_rows 与 kv_rows 数据结构，用更紧凑的行索引表示每条序列需要读取的缓存，减少无关位置参与处理。\n\n2. 哪些场景值得关注。变化直接面向多序列推理与统一缓存。发布说明还提到按序列构建缓存单元列表、忽略滑动窗口已遮蔽的位置，以及处理不等长或交错序列的微批次。对同时服务多个会话的本地服务，这是值得做同机对照的改动；对只有一个短请求的桌面聊天，不能据此宣布同样收益。\n\n3. 哪些数字不能照搬。关联 PR 在更早时间已经给出开发者测试，本次合并并没有把旧表自动变成今天的新跑分。本文不转载旧速度表，不把上游作者的单机测试称为社区独立复现，也不将后端支持的描述扩大成所有显卡均已验收。\n\n部署判断很简单：保留旧构建，固定模型权重、KV 精度、上下文、并发数和提示词，再分别比较首 token 延迟、预填充吞吐及每流生成速度。单流与总吞吐分开记录，不能只挑最快一列。当前确认的是改动已经进入新构建；你的机器能获得多少收益，仍须按实际后端重新测量。\n\n来源：[b12059 发布页](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12059)；[上游 PR #29510](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F29510)。","llama.cpp released b12059 within this reporting window, at 22:41 UTC on November 2, 2026. It incorporates changes to attention over a unified KV cache when serving multiple sequences. This is an engine release, not a model launch or an independent hardware reproduction by AraoAI.\n\n1. The problem: with multiple sequences sharing a unified cache, attention must exclude tokens belonging to other sequences. The upstream PR describes an extension, flash_attn_ext_rows, and a kv_rows structure that compactly represents relevant cache rows rather than relying on filtering unrelated positions.\n\n2. The scope: the release notes discuss per-sequence cell lists, omitting cells hidden by sliding-window attention, and handling uneven or interleaved microbatches. These changes are relevant to multi-session local servers. They do not establish the same benefit for a single short desktop chat request, nor do they prove every backend has been validated.\n\n3. The evidence boundary: the linked PR contains earlier developer measurements. Merging it today does not turn those tables into new measurements from this window. We do not reproduce those speeds as fresh community evidence or claim independent validation.\n\nKeep the old build for rollback. Compare both builds on identical weights, KV precision, context, concurrency and prompts. Record time to first token, prefill throughput and per-stream decode separately from aggregate throughput. What is verified here is inclusion in the new release; the benefit on a particular machine still needs a backend-specific measurement.\n\nSources: [b12059 release](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12059); [upstream PR #29510](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F29510).","llama.cpp GitHub release \u002F upstream PR","https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12059","2026-11-02T23:04:48.718Z"]