[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fretaz31k04nq":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},"strix-halo-mmap-lazy-mode-report-2026-11-10","RUNTIME","2026-11-09T23:05:40.885Z","Strix Halo 新故障报告：显式 mmap 与自动 lazy mode 可能错配","New Strix Halo report: explicit mmap can conflict with automatic lazy mode","11月9日新帖 #31943 给出同机对照，关联修复 #31957 尚未合并。短提示 prefill 与长请求的差别提醒：不能把局部提速倍数当成整体吞吐。","Issue #31943, filed November 9, reports a same-machine comparison; proposed fix #31957 remains unmerged. Short-prompt prefill and long requests show why local ratios are not overall throughput.",[],[],"11 月 9 日，llama.cpp 新增 #31943：作者在 Strix Halo 128 GB、Radeon 8060S、RADV\u002FVulkan 上使用 Qwen3.8-Flash-Next IQ4_NL，报告显式 mmap 与自动 lazy mode 的错配。当前只有这位报告者的一组对照，关联修复 #31957 仍开放、未合并；新发帖日期不能替代实际测量日期，本站未独立复现。\n\n1. auto 不是固定配置。原文解释，代码按设备能力把 lazy mode 自动回退为关闭，却没有区分用户是否已显式选择 mmap。这会影响 PLE 行预取路径。作者提出把回退条件改为 mmap 实际关闭时才触发；这是待审实现，不是已发布修复，也不能写成所有核显的共同故障。\n\n2. 必须看同一测试的两行。基础构建为 0c6dca2fa，原文 llama-bench 使用 pp512、n=0、重复三次：mmap\u002Fauto 为 247.09±12.71 tok\u002Fs，mmap\u002Fon 为 586.17±9.45。它测的是短提示 prefill，不是 decode。相同帖子里，默认 auto\u002Fauto 的短提示结果反而更高，不能把“手动打开”写成无条件推荐。\n\n3. 短测试不能外推长请求。单次冷启动的 30,187-token 服务端提示，对应 mmap 与 mmap\u002Fon 为 454 和 493 tok\u002Fs，差异明显小于短提示对照，且只有一次运行。MemAvailable 是主机可用内存，不是显存占用；95 GiB 是帖子所述权重规模，也不能填入实测显存列。关联 PR 还保留了一次异常慢的重复，没有把它藏掉。\n\n部署判断：出现回退或吞吐变化时，先核对构建、实际加载方式、lazy mode、提示长度与输出正确性，再做相邻 A\u002FB。不要从单机报告推导显卡排名，不要把待审补丁作为正式版本装法。本次记录只作为候选保留：缺少正式导入所需的操作系统与 decode，未写入速度库，也不计本站复现。\n\n报告：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fissues\u002F31943\n待审修复：https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F31957","Issue #31943 was filed in llama.cpp on November 9. The author reports a mismatch between explicit mmap loading and automatic lazy mode on a 128 GB Strix Halo with Radeon 8060S, RADV\u002FVulkan, running Qwen3.8-Flash-Next IQ4_NL. This is one reporter's comparison, not AraoAI replication. Fix #31957 remains open and unmerged. Filing time is not a substitute for a stated measurement time.\n\n1. Auto is not a fixed configuration. The report says a device-capability check disables lazy mode even when mmap is explicitly requested, affecting PLE row prefetch. The proposed patch limits the fallback to cases where mmap is actually off. It is not a released fix or evidence against every integrated GPU.\n\n2. Compare rows from the same test. On base commit 0c6dca2fa, llama-bench uses pp512, n=0 and three repetitions. mmap\u002Fauto reports 247.09±12.71 tok\u002Fs, versus 586.17±9.45 for mmap\u002Fon. This is short-prompt prefill, not decode. The default auto\u002Fauto row is faster still, so manual enablement is not a universal recommendation.\n\n3. A microbenchmark is not a long request. A single cold 30,187-token server prompt reports 454 versus 493 tok\u002Fs for mmap versus mmap\u002Fon. That is a different workload with only one run. MemAvailable is host available memory, not measured VRAM use; the cited 95 GiB model size is not VRAM occupancy either. The proposed fix also retains an unusually slow repetition rather than silently discarding it.\n\nFor diagnosis, pin the build, loading path, lazy mode and prompt length, confirm correct output, then run adjacent comparisons. Do not turn one machine into a hardware ranking or present an unmerged patch as a stable release. This remains a local candidate, not an imported benchmark: required operating-system and decode fields are missing.\n\nReport: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fissues\u002F31943\nUnmerged fix: https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fpull\u002F31957","llama.cpp 社区 issue 与待审 PR","https:\u002F\u002Fgithub.com\u002Fggml-org\u002Fllama.cpp\u002Fissues\u002F31943","2026-11-09T23:05:40.886Z"]