[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2wahybonxytf4":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-b12385-qwen25vl-nvfp4-scales-2026-11-15","RUNTIME","2026-11-14T23:07:06.781Z","llama.cpp修复Qwen2.5-VL的NVFP4缩放遗漏：能加载不等于算对","llama.cpp fixes missing Qwen2.5-VL NVFP4 scales: loading is not correctness","b12385纳入FFN缩放修复。贡献者报告旧图产生异常输出，本站未独立复现；PPL诊断不能替代通用能力评测。","Build b12385 includes an FFN-scale fix. The contributor reports invalid outputs in the old graph. AraoAI has not reproduced it; perplexity diagnostics are not a general ability score.",[],[],"2026年11月14日18:02 UTC，llama.cpp合并PR #32241，并在18:48 UTC发布的b12385中列入这项修复。它处理的不是“量化模型稍微笨一点”，而是Qwen2-VL架构计算图没有应用已经加载的NVFP4逐张量缩放，模型能启动，却输出异常。以下是上游变更与贡献者自报，本站没有独立复现。\n\n第一，问题出在计算路径。PR说明，build_ffn的三个缩放参数此前传入NULL；新代码传递ffn_up_s、ffn_gate_s和ffn_down_s。没有这些缩放张量的模型仍使用空值，不能把本次修复泛化成所有Qwen视觉模型的收益，更不能把其他架构也写成已经修好。\n\n第二，诊断数字要带条件。贡献者在GB10上，将nvidia\u002FQwen2.5-VL-7B-Instruct-NVFP4转换后，使用wikitext-2、上下文512、batch与ubatch均512、5个chunks和-fit off。旧基线055c3c26e的CUDA PPL为1,043,694，修复分支为9.3759；CPU从各块NaN变为9.1177。这里比较的是同一问题的数值诊断，不是公开能力榜单，也没有速度提升结论。\n\n第三，排错顺序应先正确、后性能。遇到量化权重能加载却乱码，先保存文件哈希、转换命令、引擎提交、后端和一份固定输入，再排查模板与计算图，不应立即断言“NVFP4不可用”或为了速度继续调GPU层数。修复后仍须分别检查文本、图像与实际任务输出。\n\n这项变更证明特定修复已进入构建，不证明所有后端或量化档都经过验证。部署记录必须锁定实际二进制版本，不能只写“最新版”，更不能把一组PPL恢复正常翻译成“能力完全无损”。\n\n来源：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F32241\n发布：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12385","On November 14, 2026, llama.cpp merged PR #32241 at 18:02 UTC and included it in build b12385, released at 18:48 UTC. This is not a claim that quantization merely makes a model slightly less capable. The Qwen2-VL graph had failed to apply NVFP4 per-tensor scales that the loader already loaded, allowing a model to start but produce invalid output. These are upstream changes and contributor-reported tests, not an AraoAI reproduction.\n\nThe fix passes ffn_up_s, ffn_gate_s and ffn_down_s to build_ffn instead of NULL. Models without these scales retain null values. The PR changes and tests qwen2vl only; it does not establish that other architectures with similar patterns are fixed or that all Qwen vision models gain performance.\n\nOn GB10, the contributor converted nvidia\u002FQwen2.5-VL-7B-Instruct-NVFP4 and tested wikitext-2 with context 512, batch and ubatch 512, five chunks and -fit off. Against baseline 055c3c26e, reported CUDA perplexity changed from 1,043,694 to 9.3759; CPU changed from NaN in every chunk to 9.1177. This is a targeted numerical diagnostic, not an ability leaderboard or a throughput improvement.\n\nFor deployment debugging, correctness comes before optimization. Preserve weight hashes, conversion commands, engine commits, backend settings and a fixed input before changing GPU placement or blaming the precision format. After upgrading, check text, image and real task outputs separately. A fix appearing in a build does not validate every backend or quantization. Pin the actual binary, not “latest,” and do not translate recovered perplexity into a promise of lossless capability.\n\nSource: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F32241\nRelease: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Freleases\u002Ftag\u002Fb12385","llama.cpp PR与构建发布说明","https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F32241","2026-11-14T23:07:06.782Z"]