[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f7vhcdufuiad9":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-v090-fp8-audio-support-2026-11-05","RUNTIME","2026-11-04T23:05:17.038Z","llama.cpp v0.9.0 发布：FP8 类型与音频接口扩展，不等于所有设备自动提速","llama.cpp v0.9.0 ships FP8 types and expanded audio interfaces, not universal speedups","本次窗口内发布的 v0.9.0 汇总 FP8 E4M3、多模态和后端更新。版本发布是真正的新事件，所含功能并非都在当天首次合并；支持格式、后端可用与实际收益仍须分别核验。","Version 0.9.0 was released within this window, grouping FP8 E4M3, multimodal and backend updates. The release is new, but not every included feature was merged today. Format support, backend availability and workload gains remain separate checks.",[],[],"llama.cpp 于 11 月 4 日 13:08 UTC 发布 v0.9.0，对应发布页列出的 nightly 构建为 b12110。这次新增的是一个正式版本节点，不是所有变更都在今天首次出现。对于本地部署，值得先看格式和接口的适用边界，而不是直接把更新说明当成速度承诺。\n\n1. FP8 支持具体到类型。发布说明列出 OCP FP8 E4M3 的饱和模式、转换辅助与新的模型文件类型。关联 PR 也明确把 KV cache、CUDA 和 Vulkan 工作拆在不同变更里。因此，“能识别一种 FP8 表示”与“你的后端已有完整高效路径”不能画等号，更不能把不同 FP8、NVFP4 和 MXFP4 混成同一种精度。\n\n2. 音频接口扩展有可核对的落点。版本纳入 Chatterbox TTS、Nemotron 3.5 Transcribe 及 Pocket TTS 预定义声音支持，并列出相关 mtmd API。服务端 GET \u002Fprops 可报告模型预定义声音，TTS 端点接受对应名称。它说明接口能力进入版本，不代表所有音频模型使用同一调用方式，也不替模型权重或商用许可证作保证。\n\n3. 升级要同时固定运行条件。发布页还列出 LFM3 NVFP4 GGUF、张量拆分和多个后端优化，但没有给出适用于所有设备的统一收益。本次没有 AraoAI 独立实测，本文不搬用旧 PR 的吞吐表作为当天新跑分，也不把预填充优化解释成每个 token 的生成都会同比加速。\n\n已有可用环境，先保留旧二进制、启动参数和模型文件身份，再以同一提示词及输出长度检查加载、输出正确性、首 token 延迟和生成速度。需要音频功能的用户还应核对声音列表及输入输出约定。版本号只证明用了哪份实现，不能代替该配置跑通的证据。\n\n来源：[v0.9.0 发布页](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fv0.9.0)；[FP8 PR #27512](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F27512)。","llama.cpp released v0.9.0 at 13:08 UTC on November 4. Its release page identifies b12110 as the corresponding nightly build. This is a new version milestone, not evidence that every included change first appeared today. For local deployment, the boundaries of formats and interfaces matter more than treating release notes as a speed promise.\n\n1. FP8 support refers to a specific representation. The notes list OCP FP8 E4M3 in saturating mode, conversion helpers and a new model file type. The linked PR explicitly separates KV cache, CUDA and Vulkan work into other changes. Recognizing a format does not establish a complete, efficient path on every backend. FP8, NVFP4 and MXFP4 must not be treated as interchangeable precisions.\n\n2. Audio interfaces have concrete additions. The release includes Chatterbox TTS, Nemotron 3.5 Transcribe and predefined Pocket TTS voices, alongside related mtmd APIs. GET \u002Fprops can report a model’s predefined voices and the TTS endpoint accepts their names. This confirms interfaces in the release, not a universal invocation scheme for audio models or permission to use all model weights commercially.\n\n3. Pin the environment when upgrading. The notes also list LFM3 NVFP4 GGUF support, tensor splitting and backend optimizations, but establish no universal hardware gain. AraoAI has not independently benchmarked this build. Older PR tables are not fresh measurements, and prefill improvements do not imply proportional gains in token generation.\n\nKeep the previous binary, arguments and model identity. Check loading, output correctness, time to first token and generation speed with fixed prompts and output lengths. Audio users should also check voice enumeration and input\u002Foutput contracts. A version identifies the implementation; it does not prove that a configuration works.\n\nSources: [release](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fv0.9.0); [FP8 PR](https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F27512).","ggml-org\u002Fllama.cpp 官方发布","https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fv0.9.0","2026-11-04T23:05:17.039Z"]