[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1n5jht0hdqyse":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-b12022-responses-extra-params-2026-11-01","RUNTIME","2026-10-31T23:05:53.180Z","llama.cpp b12022 修复 Responses 参数透传：接口兼容不等于参数生效","llama.cpp b12022 forwards extra_params through Responses","此前24小时内合并的修复让 \u002Fv1\u002Fresponses 接受 extra_params，可向服务端传递 seed、min_p 等设置。尚无本站端到端复现，不应把补丁发布当成全部客户端已兼容。","The merged fix forwards extra_params such as seed and min_p through \u002Fv1\u002Fresponses. AraoAI has not independently reproduced the end-to-end behavior.",[],[],"本地服务能返回文本，不代表客户端传入的采样参数真正生效。llama.cpp 在本次采集窗口内合并 PR #30959，并在 b12022 发布说明中列出这项修复：让 \u002Fv1\u002Fresponses 接受 extra_params，把 seed、min_p 等设置继续传给服务端。这是接口行为修正，不是模型更新，也不是速度提升公告。\n\n一、确认改变发生在哪层。PR 创建于10月22日，但合并时间为10月31日18:08 UTC，b12022 发布于同日18:59 UTC，两项事件均在本次窗口内。新闻依据是本次合并与发布，而非把旧提案写成当天出现的新功能。作者同时说明文档已更新，本次未加入测试；这里不能进一步宣称回归测试已经覆盖所有客户端。\n\n二、把兼容拆开验。路径存在、请求被接受、参数被转发、推理后端实际使用参数，是不同的检查。排查随机种子或采样设置不生效时，应保存完整请求体、服务端版本和日志，再检查客户端是否把扩展字段放进正确位置。只看到 HTTP 成功状态或一段正常回答，不能证明配置已经落地。\n\n三、升级只改变证据起点。已有依赖 Responses 的本地工具，可在隔离环境比较升级前后的相同请求；固定模型、模板和其他采样项，再核对参数处理。不要因为支持这个端点，就推断它与 Chat Completions 的所有扩展参数完全等价，也不要把固定 seed 理解为跨版本输出逐字一致的保证。\n\n本站目前核实的是上游合并、补丁说明和发布条目，尚未做端到端实机复现，没有性能数字或能力保留结论。部署选型时，接口名是入口，具体参数的行为才是验收对象。\n\n来源：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F30959\n发布：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12022","A local endpoint returning text does not prove that every sampling parameter reached inference. Within this collection window, llama.cpp merged PR #30959 and listed the fix in b12022: \u002Fv1\u002Fresponses now honors extra_params, allowing settings such as seed and min_p to be forwarded to the server. This is an API behavior fix, not a model release or a performance claim.\n\n1. Separate proposal time from delivery. The PR was opened on October 22, merged on October 31 at 18:08 UTC, and listed in b12022 at 18:59 UTC. The merge and release are the new events. The author says documentation was updated and tests were not added in this PR; that is not evidence of comprehensive client regression coverage.\n\n2. Verify each boundary. An existing endpoint, an accepted request, a forwarded parameter, and a backend honoring that parameter are separate checks. Preserve the full request, server version and logs when diagnosing sampling behavior. HTTP success and a plausible answer alone do not establish that configuration took effect.\n\n3. Upgrade in isolation. Compare identical requests with the same model, template and remaining sampling settings. Responses support does not imply complete parity with every Chat Completions extension. A fixed seed is not a guarantee of byte-identical output across versions.\n\nAraoAI verified the upstream merge and release, but has not independently reproduced this path on hardware. No speed or capability claim is made. The endpoint name is an entry point; parameter behavior is the acceptance criterion.\n\nSource: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F30959\nRelease: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12022","llama.cpp GitHub PR \u002F release","https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F30959","2026-10-31T23:05:53.181Z"]