[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f39fji8oxbw02s":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":15,"bodyZh":16,"bodyEn":17,"sourceName":18,"sourceUrl":19,"updatedAt":20},"llamacpp-b12611-inkling-loop-nvfp4-scale-fix-2026-11-22","NEWS","2026-11-21T23:04:43.609Z","llama.cpp b12611修复Inkling 2循环层缩放：能加载，不代表量化计算正确","llama.cpp b12611 fixes Inkling 2 loop scales: loading is not correctness","11月21日合入的#32704将循环层权重共享移到缩放张量挂接之后。影响带对应scale的NVFP4路径，不应扩大成所有GGUF都损坏；站方尚未独立复现。","PR #32704, merged November 21, re-shares loop slots after scale tensors are attached. It targets the scaled NVFP4 path, not every GGUF. AraoAI has not independently reproduced it.",[12],{"id":13,"name":14},100,"Inkling-2-Mini",[],"llama.cpp在11月21日合入PR #32704，随后发布包含该修复的b12611。它修的不是模型装不下，也不是把上下文调短，而是Inkling 2循环层在加载权重时漏接缩放张量的问题。模型能启动、能输出文字，仍然可能算错；这条更新应先作为正确性修复阅读，不作为速度宣传。\n\n1. 循环共享要等缩放挂好。按PR说明，旧的inkling2_loop在通用加载流程挂接.scale张量之前，就把物理层复制到循环槽位。额外槽位因此缺少对应scale，核心层第二、第三遍执行NVFP4专家时没有应用这些缩放。修复把共享动作放到scale处理之后，避免只复制了权重、没带上后续附加信息。\n\n2. 影响范围不能从一种量化扩成全部GGUF。作者明确说普通ggml量化不受这项问题影响。评论区复测分别给出未变化的IQ4_NL结果，以及Mainline_Support目录UD-IQ1_S修复前后的结果；这两组不能混成同一个权重的对照。本文不拿不同文件的困惑度排优劣，也不把数值诊断换算成通用能力保留率。\n\n3. 修复正确性，不等于修复所有部署问题。同一讨论仍有用户反馈Vulkan计算缓冲区OOM，并称缩小微批量可暂避；那只是单份社区反馈，不是本修复的统一部署建议。内存预算、后端支持和输出正确性需要分别验收，不能因为这个PR已合入，就推断所有卡型的长上下文也已跑通。\n\n本次核对了PR正文、窗口内更正评论、合入时间和b12611发布说明。尚无本站真机独立复现。排查时应同时保留引擎构建、权重文件和量化来源，先确认修复是否进入实际运行的二进制，再用相同输入复核输出。单次成功聊天只能说明程序跑起来，不能替代正确性验证。\n\n来源：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F32704\n发布：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12611","On November 21, llama.cpp merged PR #32704 and subsequently published release b12611 containing the fix. This is an Inkling 2 loading-correctness change, not a claim about fitting larger contexts or generating faster. A model may start and return readable text while executing part of its computation incorrectly.\n\n1. Shared loop slots must include scales. According to the PR, inkling2_loop previously copied physical layers into their loop slots before the generic loading pass attached optional .scale tensors. The additional slots therefore lacked the per-expert scales needed on the second and third passes through the core layers. The fix re-shares the slots after the scale pass.\n\n2. Do not generalize the scope to every GGUF. The author explicitly distinguishes ordinary ggml quantizations, which are unaffected by this issue. Follow-up comments report an unchanged IQ4_NL result and separate before\u002Fafter results for a Mainline_Support UD-IQ1_S file. Those are different weight files, not a single interchangeable comparison. Perplexity is a numerical diagnostic, not a universal measure of retained task capability.\n\n3. Correctness and memory failures are separate. A participant still reports Vulkan compute-buffer OOMs and a smaller microbatch workaround. That is one community report, not a universal recommendation or proof that every platform now works. Memory admission, backend support and output correctness require separate checks.\n\nAraoAI checked the PR, its in-window corrections, merge timestamp and b12611 release notes, but has not reproduced the change on hardware. Preserve the engine build, exact weight file and publisher when troubleshooting. Verify that the actual binary contains the fix before comparing identical inputs; one successful chat is not a correctness evaluation.\n\nSource: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F32704\nRelease: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12611","llama.cpp GitHub PR与发布说明","https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F32704","2026-11-21T23:04:43.610Z"]