[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f8kp230y0srjt":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-v1-api-fp8-kv-upgrade-boundaries-2026-11-14","RUNTIME","2026-11-13T23:10:00.803Z","llama.cpp 发布 v1.0.0：升级先查API与服务边界，不把版本号当速度保证","llama.cpp v1.0.0: check API and server boundaries before treating the release as a speed upgrade","正式版移除旧公共API，加入FP8 KV缓存和MoE主机缓存参数，并收紧服务安全边界。这是发布事实，不是本站性能复现。","The release removes deprecated APIs, adds FP8 KV caches and MoE host-cache controls, and tightens server security. These are release facts, not an AraoAI performance reproduction.",[],[],"llama.cpp 于2026年11月13日10:20 UTC发布 v1.0.0，落在本站本次24小时观察窗内。对本地部署者，更重要的不是“终于到1.0”，而是自己的绑定、启动脚本和服务访问方式是否仍兼容。以下依据官方发布说明整理，本站没有运行新版本性能对照。\n\n第一，接入层要重新编译和检查。发布说明列出移除35个已弃用函数，并把NextN／投机解码接口提升到 llama.h；旧接口使用者不能只替换动态库就假定兼容。直接使用命令行和自己封装C接口，是两种不同的升级风险。\n\n第二，权重量化和KV精度要分开记录。新版本加入OCP FP8 KV缓存，参数为 -ctk\u002F-ctv f8_e4m3，并支持静态缩放。这改变的是注意力缓存的存储与计算条件，不等于把模型权重换成FP8，更不能据此承诺同一张卡必然更快、能力完全不变。显存、长上下文任务表现和速度需分别对照。\n\n第三，MoE主机缓存增加CPU分担等控制；这类跨显存、内存运行的配置，不能拿“总权重除显存带宽”直接裁定速度真假。比较时至少固定模型文件、引擎构建、缓存策略、上下文、投机设置和并发，避免把新功能、不同设置及硬件变化混成一个加速结论。\n\n第四，服务端收紧工具调用、私网媒体地址、CORS及DNS重绑定等边界。原有局域网集成若升级后被拒绝，应先核对安全选项，而不是为了恢复旧行为一次性放开所有限制。保留旧二进制和启动参数，在隔离端口验收文本、图像及工具链后再切换。正式发布说明证明功能已纳入该版本，不证明每台机器上的收益。\n\n来源：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fv1.0.0","llama.cpp published v1.0.0 at 10:20 UTC on November 13, 2026, inside this reporting window. For local deployments, compatibility matters more than the milestone label. This report summarizes upstream release notes; AraoAI has not benchmarked the new release.\n\nFirst, check bindings and rebuild integrations. The notes list removal of 35 deprecated functions and promotion of the NextN\u002Fspeculative API into llama.h. A program using the C API should not assume that replacing a library is a sufficient upgrade test. CLI-only users and applications with native bindings face different risks.\n\nSecond, record weight precision and KV precision separately. OCP FP8 KV caches use -ctk\u002F-ctv f8_e4m3 and support optional static scales. That changes attention-cache storage and computation, not the precision of model weights. It does not establish a universal speed gain or unchanged task performance. Measure memory, long-context behavior and throughput separately.\n\nThird, the release adds host-memory MoE cache controls including CPU share. A deployment spanning VRAM and host RAM does not satisfy the assumptions of a simple total-weight\u002FVRAM-bandwidth bound. Pin weights, engine build, cache policy, context, speculative settings and concurrency before comparing results.\n\nFourth, server changes tighten tool execution, private media URLs, CORS and DNS-rebinding boundaries. If an existing LAN integration stops working, inspect the relevant security options rather than disabling every restriction. Preserve the old binary and command line, validate text, images and tools on an isolated port, and only then switch. Release notes establish included functionality, not a measured benefit on every machine.\n\nSource: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fv1.0.0","llama.cpp 官方发布说明","https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fv1.0.0","2026-11-13T23:10:00.804Z"]