[{"data":1,"prerenderedAt":3092},["ShallowReactive",2],{"site-info":3,"post-macaq-one-layer-at-a-time":99,"related-zh-TW-macaq-one-layer-at-a-time":1806},{"id":4,"company":5,"extension":16,"meta":17,"services":18,"social":96,"stem":97,"__hash__":98},"site\u002Fsite.yml",{"name":6,"nameEn":7,"email":8,"address":9,"addressEn":10,"lineUrl":11,"lineId":12,"duns":13,"youtube":14,"youtubeId":15},"銡策數位有限公司","Jecision Digital Co., Ltd.","contact@jecision.com","231 新北市新店區北新路二段 262 號 3 樓之 3","3F.-3, No. 262, Sec. 2, Beixin Rd., Xindian Dist., New Taipei City 231, Taiwan","https:\u002F\u002Fline.me\u002FR\u002Fti\u002Fp\u002F@374qnumi","@374qnumi","963905705","https:\u002F\u002Fwww.youtube.com\u002Fshorts\u002F6UE0WTmyjsg","6UE0WTmyjsg","yml",{},[19,43,69],{"key":20,"zh":21,"en":32},"audit",{"title":22,"tagline":23,"summary":24,"points":25,"deliverable":29,"detail":30,"duration":31},"企業 AI 導入健檢","先看清楚，再花錢","60 分鐘線上訪談，盤點最耗時、最依賴人力的營運環節；我們會告訴你哪些工作現在就能交給 AI、哪些還不成熟。",[26,27,28],"現況盤點：梳理日常營運中最耗時、最依賴人力的環節","痛點診斷：判斷哪些工作現在就能交給 AI，哪些建議先保留人力","可行性評估：依你的業務特性與預算，篩選適合的模型與工具","《企業 AI 導入建議報告書》（PDF）","含 1–2 個「開發成本最低、成效最明顯」的優先項目、技術棧建議、預期節省的人力時間與下一步","訪談 60 分鐘，報告書 3 個工作天內交付",{"title":33,"tagline":34,"summary":35,"points":36,"deliverable":40,"detail":41,"duration":42},"AI Readiness Audit","Look before you spend","A 60-minute remote audit of the workflows that eat the most time and people-power — and an honest answer on what AI can take over today, and what it cannot.",[37,38,39],"Audit the daily workflows that consume the most time and headcount","Triage which tasks AI can own today, and which should stay human","Shortlist the models and tools that fit your budget and constraints","AI Adoption Recommendation Report (PDF)","One or two highest-ROI entry points, a recommended tech stack, expected hours saved, and the concrete next step","60-minute call, report delivered within 3 business days",{"key":44,"zh":45,"en":56},"prompt",{"title":46,"tagline":47,"summary":48,"points":49,"deliverable":53,"detail":54,"duration":55},"提示詞工程","讓 AI 穩定產出，而不是碰運氣","針對你的實際任務設計提示詞與多步驟工作流，並做極限測試，確保換了資料、換了人操作也不會走鐘。",[50,51,52],"客製提示詞：ChatGPT、Claude、Gemini、Midjourney 等主流工具的專屬模板","除錯與優化：修掉幻覺、格式跑掉、時好時壞的問題","工作流建置：把複雜任務拆成多步驟提示詞與可重複的 SOP","提示詞文件或設定檔","附使用說明與範例，並提供 1 對 1 教學，確保你與團隊能自己維護","依任務複雜度報價，通常 3–7 個工作天",{"title":57,"tagline":58,"summary":59,"points":60,"deliverable":66,"detail":67,"duration":68},"Prompt Engineering","Make AI produce reliably, not by luck","Prompt sets and multi-step workflows designed for your actual tasks, stress-tested so they keep working when the data or the operator changes.",[61,62,64],"Custom prompts for ChatGPT, Claude, Gemini, Midjourney and friends",{"Debugging and tuning":63},"kill hallucinations, broken formats and flaky output",{"Workflow design":65},"break complex work into repeatable multi-step SOPs","Prompt library or exportable config","Ships with usage notes and examples, plus a 1:1 walkthrough so your team can maintain it","Quoted by complexity, typically 3–7 business days",{"key":70,"zh":71,"en":82},"poc",{"title":72,"tagline":73,"summary":74,"points":75,"deliverable":79,"detail":80,"duration":81},"極速 PoC","五天，看到能跑的東西","不談架構藍圖，直接做一個能動的雛形。五個工作天內交付線上可操作的智能體，用你真實的資料驗證。",[76,77,78],"Day 1：15 分鐘對焦，鎖定單一痛點與資料盤點","Day 2–4：模型串接、邏輯設計與雛形開發","Day 5：Demo 展示與 AI 擴展建議書","可線上操作的智能體雛形","交付範圍為概念驗證與展示，不含後續維運或大規模伺服器部署；要進階成完整系統會另行報價","5 個工作天",{"title":83,"tagline":84,"summary":85,"points":86,"deliverable":93,"detail":94,"duration":95},"Rapid PoC","Five days to something that runs","No architecture decks — a working prototype built in five business days, validated against your real data.",[87,89,91],{"Day 1":88},"15-minute scoping call, one pain point, data inventory",{"Day 2–4":90},"model integration, logic design, prototype build",{"Day 5":92},"live demo plus an AI expansion recommendation","A runnable online agent prototype","Proof of concept and demo scope only; ongoing operations and large-scale deployment are quoted separately","5 business days",[],"site","XTwARfVAOcrXNiKehsM6V6cdkd2HXy4d5yZVrSWrI3Q",{"zh":100,"en":1012},{"id":101,"title":102,"author":103,"body":104,"cover":993,"coverAlt":994,"date":995,"description":996,"draft":997,"extension":998,"featured":997,"lang":999,"meta":1000,"navigation":403,"ogImage":1001,"path":1002,"seo":1003,"slug":1004,"stem":1005,"tags":1006,"translationKey":1004,"updated":1010,"__hash__":1011},"blog\u002Fblog\u002Fzh-TW\u002Fmacaq-one-layer-at-a-time.md","量化 122B，一次只裝一層","Jecision",{"type":105,"value":106,"toc":983},"minimark",[107,130,134,137,140,147,150,153,156,159,163,166,169,172,175,178,181,184,187,190,193,283,286,289,292,295,298,301,304,307,310,313,319,389,392,409,417,420,423,426,430,433,436,439,442,445,452,499,502,505,509,512,515,518,521,524,527,530,633,636,639,642,645,657,660,664,667,670,673,680,718,721,724,802,805,824,831,834,858,865,873,876,883,886,889,937,940,943,968,976,979],[108,109,110,114],"blockquote",{},[111,112,113],"p",{},"數字來自 arXiv 2611.10649（v1 11-07、v2 11-11）與官方程式庫 README。作者是華為 Cloud & AI Lab；程式庫是 Apache 2.0。我們沒有獨立重跑，「我認為」的地方會標出來。MACaQ-R 的 R 是 refine。",[111,115,116,117,124,125],{},"論文：",[118,119,123],"a",{"href":120,"rel":121},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.10649",[122],"nofollow","arXiv 2611.10649","｜程式庫：",[118,126,129],{"href":127,"rel":128},"https:\u002F\u002Fgithub.com\u002Fhuawei-csl\u002FMACaQ",[122],"huawei-csl\u002FMACaQ",[131,132,133],"h2",{"id":133},"先講一個畫面",[111,135,136],{},"量化一個 122B 模型，64 分鐘、32.5 GB 顯示記憶體，一台單卡機就夠。就算你手上沒有 122B 也沒關係。同一套做法讓 MACaQ-R 能修好你已經下載的那個 GGUF。",[111,138,139],{},"這件事以前得先有「放得下 244 GB」的機器。",[111,141,142,146],{},[143,144,145],"code",{},"Qwen3.5-122B-A10B"," 的 FP16 權重是 244 GB。它是 96 GB 卡的 2.5 倍，兩張放不下，三張才勉強。",[111,148,149],{},"量化聽起來是「把模型變小」。但量化這個動作本身，得先有模型。旋轉、校準、算 Hessian，每一樣都吃記憶體。Hessian 用來估計「這個權重改一點，輸出會被影響多少」。",[111,151,152],{},"所以大模型量化卡住的地方，常常不是算力，是「你得先裝得下它」。",[111,154,155],{},"MACaQ 換了一個順序：一次只需要一層在 GPU 上。其餘的權重留在主機記憶體，用完就換下一層。",[111,157,158],{},"這篇講三件事。舊方法為什麼被記憶體卡住。它把哪個成本搬到推論時、代價多少。還有你手上那個 GGUF（llama.cpp 的模型檔格式）怎麼變好。",[131,160,162],{"id":161},"量化-122b要先裝得下-122b","量化 122B，要先裝得下 122B",[111,164,165],{},"先講為什麼量化大模型要旋轉。",[111,167,168],{},"權重和激活裡總有幾個特別大的值。壓到 4-bit 時，一個離群值會把整組的尺度撐大。其他值就被壓成同一個格子。",[111,170,171],{},"（這裡的 4-bit 是 MXFP4，一種 4-bit 浮點。每 32 個值共用一個 2 的冪次尺度。）",[111,173,174],{},"做法是先把資料旋轉一下。旋轉不改變矩陣乘法的結果，但會把離群值攤開。每個座標看起來比較像鐘形分布，格點才夠用。",[111,176,177],{},"問題在「什麼時候旋轉」。",[111,179,180],{},"QuaRot、SpinQuant、DuQuant++ 都是這一類。它們把旋轉矩陣融進相鄰層的權重。推論時幾乎不用多做，旋轉已經變成權重的一部分。這是它們最漂亮的地方。",[111,182,183],{},"代價在量化時。旋轉要跨層傳遞，你得知道這個模型的層怎麼排。attention 的輸出投影在哪、residual 從哪裡加回來。架構不同，融法就不同。",[111,185,186],{},"而且層跟層之間有了依賴，量化時就沒辦法「只看一層」。",[111,188,189],{},"把成本放在量化時還是推論時，是這個領域主要的分岔。沒有誰比較笨。",[111,191,192],{},"表 1：四種做法的旋轉放在哪",[194,195,196,218],"table",{},[197,198,199],"thead",{},[200,201,202,206,209,212,215],"tr",{},[203,204,205],"th",{},"方法",[203,207,208],{},"旋轉放哪",[203,210,211],{},"量化時跨層",[203,213,214],{},"只裝一層",[203,216,217],{},"推論成本",[219,220,221,239,253,266],"tbody",{},[200,222,223,227,230,233,236],{},[224,225,226],"td",{},"QuaRot",[224,228,229],{},"融進相鄰層權重",[224,231,232],{},"要",[224,234,235],{},"不行",[224,237,238],{},"幾乎不另做",[200,240,241,244,247,249,251],{},[224,242,243],{},"SpinQuant",[224,245,246],{},"融進相鄰層權重（旋轉可學習）",[224,248,232],{},[224,250,235],{},[224,252,238],{},[200,254,255,258,260,262,264],{},[224,256,257],{},"DuQuant++",[224,259,229],{},[224,261,232],{},[224,263,235],{},[224,265,238],{},[200,267,268,271,274,277,280],{},[224,269,270],{},"MACaQ",[224,272,273],{},"推論時線上做",[224,275,276],{},"不用",[224,278,279],{},"可以",[224,281,282],{},"多 2.3%",[111,284,285],{},"表 1 是概念對照，不是本文的實測名單。這篇只拿 DuQuant++ 與 MR-GPTQ 對照；論文另外也比了 QuaRot 與 FlatQuant。",[111,287,288],{},"它要講的只有一件事。MACaQ 不是更聰明的旋轉，是把旋轉換到另一個時間點。",[131,290,291],{"id":291},"它把旋轉留在推論時做",[111,293,294],{},"MACaQ 選了另一邊：旋轉不融進權重，留在推論時算。",[111,296,297],{},"每一層的 activation 進來，先乘上一個 per-channel 尺度。接著做 block-Hadamard 轉換。Hadamard 是快速、可逆的線性變換，負責把離群值攤平。",[111,299,300],{},"權重那邊，預先把反轉換吸收掉。所以結果等價，只是把同一份乘法從量化時搬到推論時。",[111,302,303],{},"關鍵不是省不省，是誰依賴誰。",[111,305,306],{},"旋轉留線上之後，每一層只需要自己的權重、自己的 activation 統計。層與層之間沒有依賴了。",[111,308,309],{},"於是量化可以串流。搬一層進 GPU、算完、寫出去、釋放，再搬下一層。同一時間，GPU 上只有一層。",[111,311,312],{},"論文的做法是 lazy loading，其餘權重留在主機記憶體。原文的說法很直白：只有正在處理的那一層需要待在 GPU 上。",[111,314,315,316,318],{},"表 2：",[143,317,145],{}," 這次的數字（論文 §Large-Model Quantization）",[194,320,321,331],{},[197,322,323],{},[200,324,325,328],{},[203,326,327],{},"項目",[203,329,330],{},"數字",[219,332,333,341,349,357,365,373,381],{},[200,334,335,338],{},[224,336,337],{},"總參數",[224,339,340],{},"122B（MoE 混合專家，A10B 是每次啟用的參數）",[200,342,343,346],{},[224,344,345],{},"FP16 權重",[224,347,348],{},"約 244 GB",[200,350,351,354],{},[224,352,353],{},"機器",[224,355,356],{},"單張 96 GB GPU（只給 504 TFLOPS FP16、1.79 TB\u002Fs，不寫型號）",[200,358,359,362],{},[224,360,361],{},"量化時 GPU 峰值",[224,363,364],{},"32.5 GB",[200,366,367,370],{},[224,368,369],{},"量化時間",[224,371,372],{},"64 分鐘（32 flips、512K 校準 token）",[200,374,375,378],{},[224,376,377],{},"WikiText-2 困惑度",[224,379,380],{},"5.14（FP16 是 5.06）",[200,382,383,386],{},[224,384,385],{},"量化後推論需求",[224,387,388],{},"74 GB，同一張卡跑得起來",[111,390,391],{},"如圖 1，左邊是「旋轉跟著層走」，量化時兩層互相依賴。右邊是「旋轉留在線上」，同一時間只有一層在 GPU 裡。",[393,394,397,398,397,405],"figure",{"className":395},[396],"article-figure","\n  ",[399,400],"video",{"src":401,"poster":402,"autoPlay":403,"loop":403,"muted":403,"playsInline":403,"ariaLabel":404},"\u002Fimages\u002Fblog\u002Fmacaq-one-layer-at-a-time\u002Fstreaming.mp4","\u002Fimages\u002Fblog\u002Fmacaq-one-layer-at-a-time\u002Fstreaming-poster.jpg",true,"左邊是旋轉跨層傳播、量化時兩層依賴；右邊是每層獨立、一次只有一層進 GPU",[406,407,408],"figcaption",{},"圖 1：左邊的轉換跟著層走，量化時要跨層處理；右邊把轉換留到推論，一層進、一層出。",[111,410,411,412,416],{},"讀者最容易誤讀的一格是 32.5 GB。那是",[413,414,415],"strong",{},"顯示記憶體","的高峰，不是全部。那 244 GB 的 FP16 權重還在主機記憶體裡，只是不用同時進 GPU。",[111,418,419],{},"所以真正的門檻是「一張夠大的卡，加一台夠大的主機」。不是「32.5 GB 就夠」。這點很容易漏。",[111,421,422],{},"還有一件事：它不改架構。同一套流程拿去跑影像模型也行。程式庫列了三個範例。FLUX.2 Klein 4B 是影像，Wan2.2 是影片，stable-audio 是音訊。論文自己只示範了一個影像模型，說法是「不需要針對架構做修改」。",[111,424,425],{},"我認為對手上什麼模型都有的人，這個「模型無關」比 64 分鐘更有價值。",[131,427,429],{"id":428},"代價多-23","代價：多 2.3%",[111,431,432],{},"線上做轉換不是免費的。相對純 MXFP4，Qwen3-32B 在 vLLM 上端到端多 2.3%。122B 那組，論文沒有量推論開銷。",[111,434,435],{},"論文另外用一張圖拆給你看。這個開銷在 decode 端量：batch 1、context 512。模型越大，占比越小，因為矩陣乘法越久，旋轉那幾個乘法相對越輕。",[111,437,438],{},"我認為這 2.3% 買到的是「不用理解架構」。這是整篇的樞紐。",[111,440,441],{},"另一組常被搞混的數字在表 3。模型是 Qwen3.8-27B，後端是 vLLM。相對 BF16，prefill 是 2.33×、decode 是 2.58×。",[111,443,444],{},"這是「相對 BF16 的吞吐量」。數字是 2.33×／2.58×，不是「相對 MXFP4 快 2.33 倍」。兩個基準差很遠，寫混了整段就報廢。",[111,446,447,448,451],{},"表 3：",[143,449,450],{},"Qwen3.8-27B"," 在 vLLM 上的品質與速度（官方 README，基準 BF16 = 1.00×）",[194,453,454,470],{},[197,455,456],{},[200,457,458,461,464,467],{},[203,459,460],{},"精度",[203,462,463],{},"困惑度 ↓（README 未寫資料集）",[203,465,466],{},"prefill",[203,468,469],{},"decode",[219,471,472,485],{},[200,473,474,477,480,483],{},[224,475,476],{},"BF16",[224,478,479],{},"2.8463",[224,481,482],{},"1.00×",[224,484,482],{},[200,486,487,490,493,496],{},[224,488,489],{},"MACaQ W4A4",[224,491,492],{},"2.8994",[224,494,495],{},"2.33×",[224,497,498],{},"2.58×",[111,500,501],{},"W4A4 的意思是權重 4-bit、激活也 4-bit。困惑度從 2.8463 到 2.8994，是 +1.9%。這不是「幾乎無損」，是「有損但很便宜」。",[111,503,504],{},"表 3 的數字來自官方 README，跟表 2 不是同一組實驗。別跨表比。",[131,506,508],{"id":507},"同樣的品質時間少一截","同樣的品質，時間少一截",[111,510,511],{},"4-bit 的格點有限。四捨五入（round-to-nearest，RTN）選最近的格子。它只看自己，不看別人。",[111,513,514],{},"MACaQ 多做一步：翻一個碼。",[111,516,517],{},"做法是把單一 FP4 碼挪到相鄰的可表示值。不是重新量化，只是換一格。",[111,519,520],{},"要不要翻，由整層的損失決定。用 Hessian 近似：這一格改了，輸出會被影響多少。然後貪心地翻，一次一個。",[111,522,523],{},"為什麼便宜？候選只有左邊那格和右邊那格。搜尋空間小，不用做連續最佳化。",[111,525,526],{},"有兩個旋鈕：翻幾輪（flips），和用多少校準 token。要快就少翻，要好就多翻。表 4 是論文附錄逐格往上爬的原始數字（Llama-3-8B）。",[111,528,529],{},"表 4：Llama-3-8B 的取捨（MXFP4 W4A4，WikiText-2 困惑度）",[194,531,532,544],{},[197,533,534],{},[200,535,536,539,541],{},[203,537,538],{},"設定",[203,540,369],{},[203,542,543],{},"困惑度 ↓",[219,545,546,557,568,579,590,601,612,622],{},[200,547,548,551,554],{},[224,549,550],{},"FP16（不量化，參考）",[224,552,553],{},"—",[224,555,556],{},"6.14",[200,558,559,562,565],{},[224,560,561],{},"RTN：只四捨五入",[224,563,564],{},"2.2 秒",[224,566,567],{},"8.29",[200,569,570,573,576],{},[224,571,572],{},"MACaQ：0 flips（旋轉＋指數篩選）",[224,574,575],{},"5.5 秒",[224,577,578],{},"7.64",[200,580,581,584,587],{},[224,582,583],{},"MACaQ：1 塊校準（2048 token）",[224,585,586],{},"14.8 秒",[224,588,589],{},"7.11",[200,591,592,595,598],{},[224,593,594],{},"MACaQ：128 flips、31 塊",[224,596,597],{},"88.1 秒",[224,599,600],{},"6.91",[200,602,603,606,609],{},[224,604,605],{},"MACaQ：512 flips、977 塊",[224,607,608],{},"636.9 秒",[224,610,611],{},"6.85",[200,613,614,617,620],{},[224,615,616],{},"MR-GPTQ：128 塊",[224,618,619],{},"319.1 秒",[224,621,589],{},[200,623,624,627,630],{},[224,625,626],{},"DuQuant++（GPTQ）：128 塊",[224,628,629],{},"624.4 秒",[224,631,632],{},"7.16",[111,634,635],{},"MACaQ 那幾列是同一條曲線上的不同取捨，校準量不一樣。MR-GPTQ 與 DuQuant++ 那兩列是各自的 128 塊設定。別逐列比速度，那是不同預算下的單點。論文自己的公平對照是 128 flips、約 20 萬 token，Llama-3-8B 上約 150 秒，落在 6.90。",[111,637,638],{},"這張表比一個「快 11×」的標題有用。你可以直接挑自己的那一列。",[111,640,641],{},"摘要的說法是：在 Llama-3-8B 上打平對手最好的困惑度，時間只要 1\u002F11。我們不換算成一個比值放進文章。表 4 是逐格往上爬的單點，摘要那條是曲線。要選型，看表。",[111,643,644],{},"如圖 2，翻一個碼的意思：四捨五入選最近的格子，翻碼是換到隔壁一格。但那次換要不要做，看的是整層的損失。",[393,646,397,648,397,654],{"className":647},[396],[649,650],"img",{"src":651,"alt":652,"loading":653},"\u002Fimages\u002Fblog\u002Fmacaq-one-layer-at-a-time\u002Fflip-grid.svg","FP4 格點示意：四捨五入到最近的點，以及翻到相鄰點的候選，取捨由整層損失決定","lazy",[406,655,656],{},"圖 2：格點之間只有鄰居可選，所以「翻一個碼」很便宜；貴的是判斷要不要翻——那要算整層的損失。示意圖，位置不是論文實測值。",[111,658,659],{},"還有一件事要講清楚：這不是「4-bit 無損」。最好的設定也是 6.14 變 6.85，困惑度多了大約 12%。RTN 是 8.29。兩者都是「有損」，差別在損失多少、以及你願不願意付時間。",[131,661,663],{"id":662},"你手上那個-gguf可以修","你手上那個 GGUF，可以修",[111,665,666],{},"上面講的是「重新量化一個模型」。多數人手上沒有那個情境，只有一堆下載好的 GGUF。",[111,668,669],{},"MACaQ-R 是給這個情境的。拿一個現成的 GGUF，配它的 FP16 原檔當 teacher。跑一次 refine。",[111,671,672],{},"碼重排，格式與 block scale 都不動。所以修完的檔案，用標準 llama.cpp 就能跑。",[111,674,675,676,679],{},"表 5：",[143,677,678],{},"Qwen3.6-35B-A3B"," 的 MACaQ-R 前後（官方 README 自測）",[194,681,682,695],{},[197,683,684],{},[200,685,686,689,692],{},[203,687,688],{},"版本",[203,690,691],{},"GGUF 大小",[203,693,694],{},"對 FP16 的 KL 散度 ↓",[219,696,697,708],{},[200,698,699,702,705],{},[224,700,701],{},"Unsloth UD-Q4_K_XL",[224,703,704],{},"22.36 GB",[224,706,707],{},"0.015731",[200,709,710,713,715],{},[224,711,712],{},"MACaQ-R Q4_K_XL",[224,714,704],{},[224,716,717],{},"0.013367",[111,719,720],{},"KL 散度是「兩個分布差多少」，越低越好。同一個大小，從 0.015731 降到 0.013367，低 15.0%。檔案沒變大，執行方式也沒變。換的只是每一格碼選哪個值。",[111,722,723],{},"指令是 README 原文，四行：",[725,726,731],"pre",{"className":727,"code":728,"language":729,"meta":730,"style":730},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight",".venv-llamacpp\u002Fbin\u002Fpython llamacpp\u002Fmacaq.py refine \\\n  --model models\u002Fmodel-q4.gguf --hf models\u002Foriginal \\\n  --teacher models\u002Foriginal-f16.gguf --calibration data\u002Fcalibration.txt \\\n  --output models\u002Fmodel-macaq-r.gguf --work-dir work\u002Frefinement\n","bash","",[143,732,733,753,770,787],{"__ignoreMap":730},[734,735,738,742,746,749],"span",{"class":736,"line":737},"line",1,[734,739,741],{"class":740},"sBMFI",".venv-llamacpp\u002Fbin\u002Fpython",[734,743,745],{"class":744},"sfazB"," llamacpp\u002Fmacaq.py",[734,747,748],{"class":744}," refine",[734,750,752],{"class":751},"sTEyZ"," \\\n",[734,754,756,759,762,765,768],{"class":736,"line":755},2,[734,757,758],{"class":744},"  --model",[734,760,761],{"class":744}," models\u002Fmodel-q4.gguf",[734,763,764],{"class":744}," --hf",[734,766,767],{"class":744}," models\u002Foriginal",[734,769,752],{"class":751},[734,771,773,776,779,782,785],{"class":736,"line":772},3,[734,774,775],{"class":744},"  --teacher",[734,777,778],{"class":744}," models\u002Foriginal-f16.gguf",[734,780,781],{"class":744}," --calibration",[734,783,784],{"class":744}," data\u002Fcalibration.txt",[734,786,752],{"class":751},[734,788,790,793,796,799],{"class":736,"line":789},4,[734,791,792],{"class":744},"  --output",[734,794,795],{"class":744}," models\u002Fmodel-macaq-r.gguf",[734,797,798],{"class":744}," --work-dir",[734,800,801],{"class":744}," work\u002Frefinement\n",[111,803,804],{},"修完直接跑，後端不用換：",[725,806,808],{"className":727,"code":807,"language":729,"meta":730,"style":730},"llama cli -m models\u002Fmodel-macaq-r.gguf\n",[143,809,810],{"__ignoreMap":730},[734,811,812,815,818,821],{"class":736,"line":737},[734,813,814],{"class":740},"llama",[734,816,817],{"class":744}," cli",[734,819,820],{"class":744}," -m",[734,822,823],{"class":744}," models\u002Fmodel-macaq-r.gguf\n",[111,825,826,827,830],{},"llama.cpp 不同版本的執行檔名稱不一樣。舊版是 ",[143,828,829],{},"llama-cli","，照你裝的那個用。",[111,832,833],{},"看起來簡單，前置條件不簡單。README 有寫的三條：",[835,836,837,849,855],"ul",{},[838,839,840,841,844,845,848],"li",{},"要 ",[413,842,843],{},"Python 3.12","，而且每個 backend 一個獨立環境。不是 ",[143,846,847],{},"pip install"," 就結束",[838,850,840,851,854],{},[413,852,853],{},"FP16 原檔","當 teacher。只有 GGUF、沒有原檔的人，這一步就卡住",[838,856,857],{},"要一份校準文字檔，以及對應的原始權重目錄",[111,859,860,861,864],{},"還有一個限制。表 5 只有一組對照，單一模型、單一 baseline。",[413,862,863],{},"不能推廣成「所有 GGUF 都能改善 15%」。"," 你的模型可能進步更多，也可能沒差。要判斷，只能自己量 KL。",[111,866,867,868,872],{},"修完之後跑多快，不是這篇的主題。那取決於位元寬與 kernel 怎麼寫。系列裡寫過：",[118,869,871],{"href":870},"\u002Fblog\u002Fquantization-kernel-dispatch","量化會不會加速，取決於你怎麼寫 kernel","。",[131,874,875],{"id":875},"別過度解讀",[111,877,878,879,882],{},"論文自己把代價寫在 Limitations 第一句。layer-local 的設計，意味著",[413,880,881],{},"跨層錯誤補償做不到","。要補，就得再加一趟跨層統計的 pass。那大概會退回其他方法的時間。",[111,884,885],{},"這是設計上的取捨，不是疏忽。",[111,887,888],{},"其他該知道的：",[835,890,891,897,903,909,915,921,931],{},[838,892,893,896],{},[413,894,895],{},"沒有第三方複現。"," 11-07 首發，對照的 DuQuant++、MR-GPTQ 都是作者自己跑的",[838,898,899,902],{},[413,900,901],{},"W4A4 的公開數字集中在少數模型。"," 別把 Qwen3.8-27B 的 +1.9% 當通則",[838,904,905,908],{},[413,906,907],{},"MoE 只量化 expert 層。"," 另外 embeddings 與 lm-head 在所有模型上都不量化",[838,910,911,914],{},[413,912,913],{},"122B 那組的預算跟其他實驗不同","：32 flips、512K 校準 token；公平對照時用的是 128 flips、約 20 萬 token",[838,916,917,920],{},[413,918,919],{},"Qwen3.5／3.6／3.8 是不同世代","，數字不要互相對照",[838,922,923,926,927,930],{},[413,924,925],{},"Ascend 不是開箱即用。"," 要 CANN 與 ",[143,928,929],{},"torch_npu","。官方只給 kernel PR，整模型還要接推理 runtime",[838,932,933,936],{},[413,934,935],{},"32.5 GB 是顯示記憶體，不是全部。"," 你的主機要有地方放那 244 GB",[111,938,939],{},"還有那個 +2.3%：它是對「純 MXFP4」算的。如果你的比較基準是 BF16，MACaQ 還是快兩倍多（表 3）。同一個方法，換一個分母，結論就換一個方向。",[131,941,942],{"id":942},"明天可以做的三件事",[944,945,946,956,962],"ol",{},[838,947,948,951,952,872],{},[413,949,950],{},"先量基準","。如果你手上已經有一個 GGUF，配 FP16 原檔算一次 KL，就有自己的一條線。README 沒有附量 KL 的腳本，得自己找。沒有基準，任何「低 15%」都是別人的數字。也先確認你手上那個不是壞的，見",[118,953,955],{"href":954},"\u002Fblog\u002Fbroken-on-arrival-quantized-models","你下載的量化模型可能是壞的",[838,957,958,961],{},[413,959,960],{},"跑一次 refine","。有原檔就跑，跑完用你平常那支 llama.cpp 跑同一題，比輸出。沒變好，你只損失一次 refine 的時間。",[838,963,964,967],{},[413,965,966],{},"要自己量化 100B 級模型的人，現在一台單卡機就夠。"," 前提是主機記憶體放得下 FP16 權重。",[111,969,970,971,975],{},"如果只是想在自己電腦上跑大模型，先看",[118,972,974],{"href":973},"\u002Fblog\u002Fstrata-125b-on-12gb","125B 模型怎麼跑在 12GB 顯卡上","。那是另一條路：不改權重，改誰在什麼時候被載入。",[111,977,978],{},"量化的門檻換了位置：從一張卡，搬到一台主機。",[980,981,982],"style",{},"html pre.shiki code .sBMFI, html code.shiki .sBMFI{--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B}html pre.shiki code .sfazB, html code.shiki .sfazB{--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D}html pre.shiki code .sTEyZ, html code.shiki .sTEyZ{--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8}html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":730,"searchDepth":755,"depth":755,"links":984},[985,986,987,988,989,990,991,992],{"id":133,"depth":755,"text":133},{"id":161,"depth":755,"text":162},{"id":291,"depth":755,"text":291},{"id":428,"depth":755,"text":429},{"id":507,"depth":755,"text":508},{"id":662,"depth":755,"text":663},{"id":875,"depth":755,"text":875},{"id":942,"depth":755,"text":942},"\u002Fimages\u002Fblog\u002Fmacaq-one-layer-at-a-time.svg","標題「量化 122B，一次只裝一層」的技術筆記封面","2026-11-19","量化一個 122B 模型只要 64 分鐘、32.5 GB 顯示記憶體——因為它不把旋轉矩陣跨層傳播，每層獨立，一次只裝一層。權重留在主機記憶體；在 32B 上量到的代價是推論多 2.3%。MACaQ-R 在一個 35B GGUF 上把 KL 修低 15%，前提是有 FP16 原檔。",false,"md","zh-TW",{},"\u002Fimages\u002Fblog\u002Fmacaq-one-layer-at-a-time.jpg","\u002Fblog\u002Fzh-tw\u002Fmacaq-one-layer-at-a-time",{"title":102,"description":996},"macaq-one-layer-at-a-time","blog\u002Fzh-TW\u002Fmacaq-one-layer-at-a-time",[1007,1008,1009],"quantization","local-llm","gguf",null,"im6BNKmEd2jNobz5OpyhsjwIzUViuMdNqM7ywhcggCI",{"id":1013,"title":1014,"author":103,"body":1015,"cover":1795,"coverAlt":1796,"date":995,"description":1797,"draft":997,"extension":998,"featured":997,"lang":1798,"meta":1799,"navigation":403,"ogImage":1800,"path":1801,"seo":1802,"slug":1004,"stem":1803,"tags":1804,"translationKey":1004,"updated":1010,"__hash__":1805},"blog\u002Fblog\u002Fen\u002Fmacaq-one-layer-at-a-time.md","Quantize a 122B Model, One Layer at a Time",{"type":105,"value":1016,"toc":1785},[1017,1032,1036,1039,1042,1047,1050,1053,1056,1059,1063,1066,1069,1072,1075,1078,1081,1084,1087,1090,1093,1171,1174,1177,1181,1184,1187,1190,1193,1196,1202,1271,1274,1285,1292,1295,1298,1301,1305,1308,1311,1314,1317,1324,1330,1368,1371,1374,1378,1381,1384,1387,1390,1393,1396,1399,1495,1498,1501,1504,1507,1517,1520,1524,1527,1530,1533,1539,1572,1575,1578,1626,1629,1643,1649,1652,1675,1682,1689,1693,1700,1703,1750,1753,1757,1780,1783],[108,1018,1019],{},[111,1020,1021,1022,1026,1027,1031],{},"Sources. The ",[118,1023,1025],{"href":120,"rel":1024},[122],"paper"," (arXiv 2611.10649, v1 Nov 7 2026, v2 Nov 11 2026; Huawei Cloud & AI Lab) supplies the 122B run, the Llama-3-8B ladder in its appendix and the limitations. The ",[118,1028,1030],{"href":127,"rel":1029},[122],"repository README"," (Apache 2.0) supplies the vLLM numbers, the MACaQ-R comparison and the command. We did not rerun anything; where a claim is ours, it says so.",[131,1033,1035],{"id":1034},"the-model-does-not-fit","The model does not fit",[111,1037,1038],{},"A 122B model, quantized in 64 minutes, peaking at 32.5 GB of GPU memory, on a single-GPU machine. That matters even if you never touch a 122B model: the same layout is what lets MACaQ-R improve a GGUF you already downloaded.",[111,1040,1041],{},"Before this, you first needed a machine that could hold 244 GB.",[111,1043,1044,1046],{},[143,1045,145],{}," weighs 244 GB in FP16. That is 2.5× a 96 GB card. Two cards hold 192 GB, so it does not fit. Three hold 288 GB, with 44 GB to spare.",[111,1048,1049],{},"Quantization sounds like making a model smaller. But quantization is an operation that needs the model first. Rotations, calibration, Hessians: each one costs memory. Hessians here mean \"how far does the output move if this weight nudges.\"",[111,1051,1052],{},"So the wall for large-model quantization is rarely arithmetic throughput. It is the simple fact that you have to hold the thing.",[111,1054,1055],{},"MACaQ reverses the order. Only one layer has to be on the GPU at a time. Everything else waits in host memory.",[111,1057,1058],{},"Four things below: why the old layout hits the memory wall, what the rotation costs once it moves, how the flips buy quality back, and how to improve the GGUF you already have.",[131,1060,1062],{"id":1061},"quantizing-122b-means-holding-122b","Quantizing 122B means holding 122B",[111,1064,1065],{},"Start with why quantization needs a rotation at all.",[111,1067,1068],{},"Every weight matrix and activation has a few large values. Squeezing to 4 bits, one outlier stretches the shared block scale. Everything else collapses onto the same grid point.",[111,1070,1071],{},"(The 4 bits here are MXFP4: a 4-bit float where 32 values share one power-of-two scale.)",[111,1073,1074],{},"The fix is to rotate the data first. A rotation does not change what the matrix product means, but it spreads the outliers out. Every coordinate looks closer to a bell curve, and the grid starts to cover it.",[111,1076,1077],{},"The question is when the rotation happens.",[111,1079,1080],{},"QuaRot, SpinQuant and DuQuant++ are one family. They fuse the rotation into the neighbouring layer's weights. At inference there is almost nothing extra to do, because the rotation is now part of the weight. That is the elegant part.",[111,1082,1083],{},"The bill arrives at quantization time. The rotation travels across layers, so you need to know how the model is wired: where the attention output projection sits, where the residual comes back in. Different architecture, different fusion.",[111,1085,1086],{},"And once layers depend on each other, you cannot quantize one at a time.",[111,1088,1089],{},"Whether you pay at quantization time or at inference is the real fork in this field.",[111,1091,1092],{},"Table 1: where each method keeps the rotation",[194,1094,1095,1114],{},[197,1096,1097],{},[200,1098,1099,1102,1105,1108,1111],{},[203,1100,1101],{},"Method",[203,1103,1104],{},"Rotation lives",[203,1106,1107],{},"Cross-layer at quant time",[203,1109,1110],{},"One layer at a time",[203,1112,1113],{},"Inference cost",[219,1115,1116,1132,1145,1157],{},[200,1117,1118,1120,1123,1126,1129],{},[224,1119,226],{},[224,1121,1122],{},"fused into the next layer's weights",[224,1124,1125],{},"yes",[224,1127,1128],{},"no",[224,1130,1131],{},"almost nothing extra",[200,1133,1134,1136,1139,1141,1143],{},[224,1135,243],{},[224,1137,1138],{},"fused, with learned rotations",[224,1140,1125],{},[224,1142,1128],{},[224,1144,1131],{},[200,1146,1147,1149,1151,1153,1155],{},[224,1148,257],{},[224,1150,1122],{},[224,1152,1125],{},[224,1154,1128],{},[224,1156,1131],{},[200,1158,1159,1161,1164,1166,1168],{},[224,1160,270],{},[224,1162,1163],{},"online, at inference",[224,1165,1128],{},[224,1167,1125],{},[224,1169,1170],{},"+2.3%",[111,1172,1173],{},"Table 1 is a conceptual map, not the paper's benchmark list. This essay compares against DuQuant++ and MR-GPTQ only; the paper also benchmarks QuaRot and FlatQuant.",[111,1175,1176],{},"The map says one thing: the rotation itself is not smarter, it just runs at a different time. The quality comes later, from the flips.",[131,1178,1180],{"id":1179},"moving-the-rotation-to-inference-time","Moving the rotation to inference time",[111,1182,1183],{},"MACaQ takes the other branch. The rotation stays out of the weights and runs at inference.",[111,1185,1186],{},"Each layer's incoming activations get a per-channel scale, then a block-Hadamard transform. Hadamard is a fast, invertible linear map whose job is to flatten outliers.",[111,1188,1189],{},"The inverse is absorbed into the weights at quantization time. The arithmetic is equivalent. The same multiply just moves from the quantizer to the runtime.",[111,1191,1192],{},"The interesting consequence is not the saving, it is who depends on whom. With the rotation online, a layer needs only its own weights and its own activation statistics. Nothing points backwards.",[111,1194,1195],{},"So quantization can stream: load a layer onto the GPU, quantize it, write it out, free it, load the next. The paper calls this streaming with lazy loading, and puts it in one sentence; the remaining weights stay in host memory, and only the layer being processed has to sit on the GPU.",[111,1197,1198,1199,1201],{},"Table 2: the ",[143,1200,145],{}," run",[194,1203,1204,1214],{},[197,1205,1206],{},[200,1207,1208,1211],{},[203,1209,1210],{},"Item",[203,1212,1213],{},"Value",[219,1215,1216,1224,1232,1240,1247,1255,1263],{},[200,1217,1218,1221],{},[224,1219,1220],{},"Parameters",[224,1222,1223],{},"122B (mixture-of-experts; about 10B active per token)",[200,1225,1226,1229],{},[224,1227,1228],{},"FP16 weights",[224,1230,1231],{},"about 244 GB",[200,1233,1234,1237],{},[224,1235,1236],{},"Machine",[224,1238,1239],{},"one 96 GB GPU (the paper lists 504 TFLOPS FP16 and 1.79 TB\u002Fs, not a model name)",[200,1241,1242,1245],{},[224,1243,1244],{},"Peak GPU memory while quantizing",[224,1246,364],{},[200,1248,1249,1252],{},[224,1250,1251],{},"Quantization time",[224,1253,1254],{},"64 minutes (32 flips, 512K calibration tokens)",[200,1256,1257,1260],{},[224,1258,1259],{},"WikiText-2 perplexity",[224,1261,1262],{},"5.14 (FP16 is 5.06)",[200,1264,1265,1268],{},[224,1266,1267],{},"Footprint of the result",[224,1269,1270],{},"74 GB; runs on the same card",[111,1272,1273],{},"Figure 1, left: the rotation travels with the layers, so quantizing two of them couples them. Right: the rotation stays online, so one layer is on the GPU at a time.",[393,1275,397,1277,397,1282],{"className":1276},[396],[399,1278],{"src":1279,"poster":1280,"autoPlay":403,"loop":403,"muted":403,"playsInline":403,"ariaLabel":1281},"\u002Fimages\u002Fblog\u002Fen\u002Fmacaq-one-layer-at-a-time\u002Fstreaming.mp4","\u002Fimages\u002Fblog\u002Fen\u002Fmacaq-one-layer-at-a-time\u002Fstreaming-poster.jpg","Left: the rotation propagates across layers and couples them during quantization. Right: each layer stands alone, one layer on the GPU at a time.",[406,1283,1284],{},"Figure 1: on the left the transform travels with the layers, so quantization has to look across them; on the right it stays at inference and each layer goes in and out on its own. The 32.5 GB badge is the 122B number from Table 2, not a general bound.",[111,1286,1287,1288,1291],{},"The number people misread is 32.5 GB. That is ",[413,1289,1290],{},"device memory"," at its peak, not the whole footprint. The 244 GB of FP16 weights is still in host memory; it is just never resident on the GPU at once.",[111,1293,1294],{},"So the real requirement is a big enough card plus a big enough host. Not \"32.5 GB is enough.\" Almost every repost drops that half.",[111,1296,1297],{},"No architecture changes, either. The repo lists FLUX.2 Klein 4B for images, Wan2.2 for video and stable-audio for audio. The paper itself demonstrates one image model and describes the method as applying \"without architecture-specific modifications.\"",[111,1299,1300],{},"If you keep a zoo of models, that portability matters more than the 64 minutes.",[131,1302,1304],{"id":1303},"the-bill-23","The bill: 2.3%",[111,1306,1307],{},"The online transform is not free. Against pure MXFP4, MACaQ adds 2.3% end-to-end on Qwen3-32B in vLLM. The paper never measures the 122B run; that is a 32B number.",[111,1309,1310],{},"The paper's Figure 3 breaks it down: the cost is measured on decode, batch 1, context 512. It shrinks as models grow, because bigger matrices take longer and a few extra multiplies weigh less.",[111,1312,1313],{},"I think what that 2.3% buys is \"you do not have to understand the architecture.\" That is the hinge of the design.",[111,1315,1316],{},"The other number that gets mangled is in Table 3: W4A4 (4-bit weights, 4-bit activations) on Qwen3.8-27B in vLLM, a serving engine. MACaQ runs at 2.33× prefill and 2.58× decode.",[111,1318,1319,1320,1323],{},"Those are throughput ratios ",[413,1321,1322],{},"against BF16",", not against pure MXFP4. The two denominators answer different questions; mixing them wrecks the paragraph.",[111,1325,1326,1327,1329],{},"Table 3: ",[143,1328,450],{}," in vLLM (README; BF16 = 1.00×)",[194,1331,1332,1346],{},[197,1333,1334],{},[200,1335,1336,1339,1342,1344],{},[203,1337,1338],{},"Precision",[203,1340,1341],{},"Perplexity ↓ (README; dataset unstated)",[203,1343,466],{},[203,1345,469],{},[219,1347,1348,1358],{},[200,1349,1350,1352,1354,1356],{},[224,1351,476],{},[224,1353,479],{},[224,1355,482],{},[224,1357,482],{},[200,1359,1360,1362,1364,1366],{},[224,1361,489],{},[224,1363,492],{},[224,1365,495],{},[224,1367,498],{},[111,1369,1370],{},"Perplexity goes from 2.8463 to 2.8994, a 1.9% increase. That is not lossless; it is lossy and cheap, which is a different claim.",[111,1372,1373],{},"Table 3 comes from the README and is a different experiment from Table 2. Do not compare across tables.",[131,1375,1377],{"id":1376},"the-rounding-is-the-easy-part","The rounding is the easy part",[111,1379,1380],{},"Four bits give you a small grid. Round-to-nearest (RTN) picks the closest point. It looks only at the weight itself.",[111,1382,1383],{},"MACaQ does one more thing: it flips a code.",[111,1385,1386],{},"A flip moves a single FP4 code to an adjacent representable value. Not a re-quantization, just a different point in the same grid.",[111,1388,1389],{},"Whether to flip is decided by the whole layer's loss. A Hessian approximation asks how far the output moves if this code changes. Then it flips greedily, one code at a time.",[111,1391,1392],{},"Why is that cheap? The only candidates are the neighbour on the left and the neighbour on the right. The search space is tiny. No continuous optimization.",[111,1394,1395],{},"Two knobs control it: how many flips, and how many calibration tokens. Fast means few flips, good means many. Table 4 is the appendix ladder on Llama-3-8B.",[111,1397,1398],{},"Table 4: the trade on Llama-3-8B (MXFP4 W4A4, WikiText-2 perplexity)",[194,1400,1401,1414],{},[197,1402,1403],{},[200,1404,1405,1408,1411],{},[203,1406,1407],{},"Setting",[203,1409,1410],{},"Time",[203,1412,1413],{},"Perplexity ↓",[219,1415,1416,1425,1435,1445,1455,1465,1475,1485],{},[200,1417,1418,1421,1423],{},[224,1419,1420],{},"FP16 (reference)",[224,1422,553],{},[224,1424,556],{},[200,1426,1427,1430,1433],{},[224,1428,1429],{},"RTN only",[224,1431,1432],{},"2.2 s",[224,1434,567],{},[200,1436,1437,1440,1443],{},[224,1438,1439],{},"MACaQ, 0 flips (rotation + exponent screen)",[224,1441,1442],{},"5.5 s",[224,1444,578],{},[200,1446,1447,1450,1453],{},[224,1448,1449],{},"MACaQ, 1 calibration block (2048 tokens)",[224,1451,1452],{},"14.8 s",[224,1454,589],{},[200,1456,1457,1460,1463],{},[224,1458,1459],{},"MACaQ, 128 flips, 31 blocks",[224,1461,1462],{},"88.1 s",[224,1464,600],{},[200,1466,1467,1470,1473],{},[224,1468,1469],{},"MACaQ, 512 flips, 977 blocks",[224,1471,1472],{},"636.9 s",[224,1474,611],{},[200,1476,1477,1480,1483],{},[224,1478,1479],{},"MR-GPTQ, 128 blocks",[224,1481,1482],{},"319.1 s",[224,1484,589],{},[200,1486,1487,1490,1493],{},[224,1488,1489],{},"DuQuant++ (GPTQ), 128 blocks",[224,1491,1492],{},"624.4 s",[224,1494,632],{},[111,1496,1497],{},"The MACaQ rows are points on one curve with different calibration budgets. The MR-GPTQ and DuQuant++ rows use 128 blocks each. Do not read down the time column and call it a race; those are different budgets. The paper's own fair comparison is 128 flips and about 200K tokens, roughly 150 s on Llama-3-8B, landing at 6.90.",[111,1499,1500],{},"This table is more useful than a headline ratio: you can pick your own row.",[111,1502,1503],{},"The abstract claims MACaQ \"already matches their best perplexity on Llama-3-8B in 11× less quantization time.\" We did not collapse that into a single ratio: Table 4 is a set of single points, the abstract is describing a curve. For a decision, read the table.",[111,1505,1506],{},"Figure 2: rounding picks the nearest point. A flip moves one code to the neighbour, and the layer's loss decides whether that happens.",[393,1508,397,1510,397,1514],{"className":1509},[396],[649,1511],{"src":1512,"alt":1513,"loading":653},"\u002Fimages\u002Fblog\u002Fen\u002Fmacaq-one-layer-at-a-time\u002Fflip-grid.svg","Two choices on an FP4 grid: round to nearest picks 1.0, a single flip moves to 1.5, and the layer loss decides",[406,1515,1516],{},"Figure 2: the only candidates are the neighbours, which makes a flip cheap; the expensive part is deciding, because that needs the layer's loss. Schematic, not measured values.",[111,1518,1519],{},"One thing to say plainly: this is not lossless 4-bit. Even the best setting goes from 6.14 to 6.85, about 12% worse perplexity. RTN is 8.29. Both are lossy. The difference is how much, and whether you will pay the time.",[131,1521,1523],{"id":1522},"fixing-the-gguf-you-already-have","Fixing the GGUF you already have",[111,1525,1526],{},"Everything above is \"quantize a model from scratch.\" Most people are not doing that. They have a folder of downloaded GGUFs.",[111,1528,1529],{},"MACaQ-R is for that case. Take an existing GGUF and its original FP16 file as a teacher, run one refine pass.",[111,1531,1532],{},"The weight codes get rewritten. The native format and block scales stay. So the result still runs on a stock llama.cpp build.",[111,1534,1535,1536,1538],{},"Table 5: ",[143,1537,678],{}," before and after MACaQ-R (README, single run)",[194,1540,1541,1554],{},[197,1542,1543],{},[200,1544,1545,1548,1551],{},[203,1546,1547],{},"Version",[203,1549,1550],{},"GGUF size",[203,1552,1553],{},"KL to FP16 ↓",[219,1555,1556,1564],{},[200,1557,1558,1560,1562],{},[224,1559,701],{},[224,1561,704],{},[224,1563,707],{},[200,1565,1566,1568,1570],{},[224,1567,712],{},[224,1569,704],{},[224,1571,717],{},[111,1573,1574],{},"KL divergence measures how far two distributions are apart; lower is better. Same size, 0.015731 down to 0.013367, a 15.0% drop. Same file size, same runtime, different code in each slot.",[111,1576,1577],{},"The command is straight from the README:",[725,1579,1580],{"className":727,"code":728,"language":729,"meta":730,"style":730},[143,1581,1582,1592,1604,1616],{"__ignoreMap":730},[734,1583,1584,1586,1588,1590],{"class":736,"line":737},[734,1585,741],{"class":740},[734,1587,745],{"class":744},[734,1589,748],{"class":744},[734,1591,752],{"class":751},[734,1593,1594,1596,1598,1600,1602],{"class":736,"line":755},[734,1595,758],{"class":744},[734,1597,761],{"class":744},[734,1599,764],{"class":744},[734,1601,767],{"class":744},[734,1603,752],{"class":751},[734,1605,1606,1608,1610,1612,1614],{"class":736,"line":772},[734,1607,775],{"class":744},[734,1609,778],{"class":744},[734,1611,781],{"class":744},[734,1613,784],{"class":744},[734,1615,752],{"class":751},[734,1617,1618,1620,1622,1624],{"class":736,"line":789},[734,1619,792],{"class":744},[734,1621,795],{"class":744},[734,1623,798],{"class":744},[734,1625,801],{"class":744},[111,1627,1628],{},"Then it runs, with nothing to swap in the backend:",[725,1630,1631],{"className":727,"code":807,"language":729,"meta":730,"style":730},[143,1632,1633],{"__ignoreMap":730},[734,1634,1635,1637,1639,1641],{"class":736,"line":737},[734,1636,814],{"class":740},[734,1638,817],{"class":744},[734,1640,820],{"class":744},[734,1642,823],{"class":744},[111,1644,1645,1646,1648],{},"llama.cpp binaries changed names across versions. Older builds use ",[143,1647,829],{},"; use whichever you installed.",[111,1650,1651],{},"Simple command, not-simple prerequisites. Three of them, from the README:",[835,1653,1654,1661,1668],{},[838,1655,1656,1658,1659],{},[413,1657,843],{},", and a separate environment per backend. Not just ",[143,1660,847],{},[838,1662,1663,1664,1667],{},"The ",[413,1665,1666],{},"original FP16 file"," as teacher. If all you have is the GGUF, you stop here",[838,1669,1670,1671,1674],{},"A ",[413,1672,1673],{},"calibration text file"," and the original weights directory",[111,1676,1677,1678,1681],{},"One more limit. Table 5 is a single comparison: one model, one baseline. ",[413,1679,1680],{},"Do not generalize it to \"any GGUF improves by 15%.\""," Yours may improve more, or not at all. Measure KL yourself.",[111,1683,1684,1685,1688],{},"How fast it runs afterwards is a separate topic, decided by bit width and kernel. We covered that in ",[118,1686,1687],{"href":870},"Quantization Speed Is Often a Kernel Problem",".",[131,1690,1692],{"id":1691},"what-this-does-not-prove","What this does not prove",[111,1694,1695,1696,1699],{},"The paper puts the cost in the first line of Limitations. A layer-local design means ",[413,1697,1698],{},"cross-layer error compensation is impossible."," A second pass over cross-layer statistics would likely bring back the slow quantization of the other methods. That is a design trade, not an oversight.",[111,1701,1702],{},"Also worth knowing:",[835,1704,1705,1711,1717,1723,1729,1735,1744],{},[838,1706,1707,1710],{},[413,1708,1709],{},"No third-party replication yet."," First posted Nov 7; the DuQuant++ and MR-GPTQ comparisons were run by the authors",[838,1712,1713,1716],{},[413,1714,1715],{},"W4A4 numbers are concentrated on a few models."," Do not read the +1.9% on Qwen3.8-27B as a general result",[838,1718,1719,1722],{},[413,1720,1721],{},"MoE: expert layers only."," Embeddings and lm-head are skipped in every model, not just MoE",[838,1724,1725,1728],{},[413,1726,1727],{},"The 122B budget differs from the rest."," 32 flips and 512K calibration tokens; the fair comparison used 128 flips and about 200K tokens",[838,1730,1731,1734],{},[413,1732,1733],{},"Qwen3.5 \u002F 3.6 \u002F 3.8 are different generations."," Do not cross-compare",[838,1736,1737,1740,1741,1743],{},[413,1738,1739],{},"Ascend is not plug and play."," Needs CANN and ",[143,1742,929],{},"; upstream has kernel PRs only, and full-model execution still needs a serving runtime",[838,1745,1746,1749],{},[413,1747,1748],{},"32.5 GB is device memory, not the whole footprint."," Your host still has to hold 244 GB",[111,1751,1752],{},"And the +2.3% again: it is measured against pure MXFP4. Against BF16 the same method is over 2× faster (Table 3). Same method, different denominator, opposite headline.",[131,1754,1756],{"id":1755},"three-things-to-do-tomorrow","Three things to do tomorrow",[944,1758,1759,1768,1774],{},[838,1760,1761,1764,1765,1688],{},[413,1762,1763],{},"Measure a baseline first."," If you own a GGUF, compute KL against its FP16 original. The README does not give you a KL script; you will have to find one. Without a baseline, \"15% lower\" is somebody else's number. Check the file is not broken first, too: ",[118,1766,1767],{"href":954},"Your quantized model might be broken",[838,1769,1770,1773],{},[413,1771,1772],{},"Run one refine pass."," If you have the original, run it, then run the same prompt on your usual llama.cpp build and compare outputs. Worst case you lose one refine.",[838,1775,1776,1779],{},[413,1777,1778],{},"Quantizing something 100B-class is a single-GPU job now."," Provided the host can hold the FP16 weights.",[111,1781,1782],{},"The gate moved. It used to be a card. Now it is a machine.",[980,1784,982],{},{"title":730,"searchDepth":755,"depth":755,"links":1786},[1787,1788,1789,1790,1791,1792,1793,1794],{"id":1034,"depth":755,"text":1035},{"id":1061,"depth":755,"text":1062},{"id":1179,"depth":755,"text":1180},{"id":1303,"depth":755,"text":1304},{"id":1376,"depth":755,"text":1377},{"id":1522,"depth":755,"text":1523},{"id":1691,"depth":755,"text":1692},{"id":1755,"depth":755,"text":1756},"\u002Fimages\u002Fblog\u002Fen\u002Fmacaq-one-layer-at-a-time.svg","Cover for the English essay, Quantize a 122B Model One Layer at a Time","MACaQ quantizes a 122B model in 64 minutes on 32.5 GB GPU memory, one layer at a time, rest in host RAM. At 32B the inference cost is 2.3%. MACaQ-R cut KL 15% on a 35B GGUF, given the FP16 original.","en",{},"\u002Fimages\u002Fblog\u002Fen\u002Fmacaq-one-layer-at-a-time.jpg","\u002Fblog\u002Fen\u002Fmacaq-one-layer-at-a-time",{"title":1014,"description":1797},"blog\u002Fen\u002Fmacaq-one-layer-at-a-time",[1007,1008,1009],"h1P4p-K3DukBf765f6-nLM5UhLte72EnDfcR2_c1VJ0",[1807,2277],{"id":1808,"title":1809,"author":103,"body":1810,"cover":2265,"coverAlt":2266,"date":2267,"description":2268,"draft":997,"extension":998,"featured":997,"lang":999,"meta":2269,"navigation":403,"ogImage":2270,"path":2271,"seo":2272,"slug":2273,"stem":2274,"tags":2275,"translationKey":2273,"updated":1010,"__hash__":2276},"blog\u002Fblog\u002Fzh-TW\u002Fbroken-on-arrival-quantized-models.md","你下載的量化模型可能是壞的，而且 checksum 驗不出來",{"type":105,"value":1811,"toc":2254},[1812,1821,1823,1830,1833,1836,1843,1847,1850,1857,1863,1874,1877,1880,1887,1897,1904,1908,1911,1969,1972,1978,1982,1985,1992,1999,2002,2006,2009,2114,2124,2135,2138,2151,2161,2175,2177,2195,2198,2203,2209,2216,2219,2241],[108,1813,1814],{},[111,1815,1816,1817,1820],{},"數字來自 arXiv 2609.05881v3 與作者的 ",[143,1818,1819],{},"quantcheck"," 稽核資料；「今天還在不在」是我們自己查的，腳本在文末。",[131,1822,133],{"id":133},[111,1824,1825,1826,1829],{},"你在筆電上 ",[143,1827,1828],{},"ollama pull qwen2.5-coder:3b","，選了一個 2-bit 的版本，因為記憶體不夠。",[111,1831,1832],{},"模型跑起來了，有回應，速度也不錯。但你叫它寫一個排序函式，它寫出來的東西編譯不過；再試一題，還是不過。",[111,1834,1835],{},"你會先懷疑自己：prompt 寫錯？溫度太高？還是這顆模型本來就這麼弱？",[111,1837,1838,1839,1842],{},"有一篇論文把這個懷疑變成了普查：",[413,1840,1841],{},"Ollama 官方庫裡 305 個「能寫程式」的模型 tag，有 7 個在 164 題 HumanEval+ 上幾乎全錯。"," 其中六個一題都沒對。",[131,1844,1846],{"id":1845},"最反直覺的地方檔案是好的","最反直覺的地方：檔案是好的",[111,1848,1849],{},"如果模型壞了，你會想「是不是下載壞了、檔案損毀」。",[111,1851,1852,1853,1856],{},"作者把其中五個檔案從原始權重重新量化回來，",[413,1854,1855],{},"tensor 對 tensor 比對，完全一致","。Qwen2.5-Coder-3B 那四個 tag，434 個 tensor 全部相同。",[111,1858,1859,1860],{},"也就是說：",[413,1861,1862],{},"這些檔案是「正確的壞檔案」。",[111,1864,1865,1866,1869,1870,1873],{},"checksum 會過、檔案大小會對、",[143,1867,1868],{},"ollama show"," 看起來一切正常——因為檔案真的沒問題。壞的是",[413,1871,1872],{},"量化的配方","：同樣的權重、同樣的工具鏈、同樣的 tensor 組合，少了校準，就把模型壓成了不會寫程式的形狀。",[131,1875,1876],{"id":1876},"為什麼會這樣",[111,1878,1879],{},"量化是「把權重從 16 bit 壓到 2–4 bit」。壓的過程要不斷決定「這個誤差要丟給哪個權重」。",[111,1881,1882,1883,1886],{},"預設的做法是均勻地丟；比較好的做法是",[413,1884,1885],{},"看校準資料","，把誤差集中在不重要的權重上——llama.cpp 的 importance matrix（imatrix）就是做這件事。",[111,1888,1889,1890,1893,1894,872],{},"沒有 imatrix，小模型在低 bit 特別容易崩。這次的七個 suspects 全部落在 ",[413,1891,1892],{},"q2_K 到 q3_K_L、模型 ≤4B"," 這個範圍；同一個範圍有 48 個 tag，等於 ",[413,1895,1896],{},"14.6%",[111,1898,1899,1900,1903],{},"但要注意：",[413,1901,1902],{},"配方本身不能預測成敗","。同樣的無校準配方，Qwen2.5-Coder-1.5B 的 q2_K–q3_K_L 全部正常（13–15 題）。所以不是「低 bit 一定壞」，是「壞了你不會知道」。",[131,1905,1907],{"id":1906},"加回校準救回大部分","加回校準，救回大部分",[111,1909,1910],{},"作者固定權重、工具鏈與 tensor 組合，只補上 imatrix：",[194,1912,1913,1926],{},[197,1914,1915],{},[200,1916,1917,1920,1923],{},[203,1918,1919],{},"Tag",[203,1921,1922],{},"原本",[203,1924,1925],{},"補 imatrix 後（滿分 164）",[219,1927,1928,1939,1949,1959],{},[200,1929,1930,1933,1936],{},[224,1931,1932],{},"Qwen2.5-Coder-3B q2_K",[224,1934,1935],{},"0",[224,1937,1938],{},"106–113",[200,1940,1941,1944,1946],{},[224,1942,1943],{},"Qwen2.5-Coder-3B q3_K_M",[224,1945,1935],{},[224,1947,1948],{},"76–89",[200,1950,1951,1954,1956],{},[224,1952,1953],{},"Qwen2.5-Coder-3B q3_K_L",[224,1955,1935],{},[224,1957,1958],{},"80–86",[200,1960,1961,1964,1966],{},[224,1962,1963],{},"Qwen2.5-Coder-3B q3_K_S",[224,1965,1935],{},[224,1967,1968],{},"13–30（救不回）",[111,1970,1971],{},"（論文摘要寫「0 → 76–113」；上面是 v3.4 稽核資料的分級數字，我以資料為準。）",[111,1973,1974,1977],{},[413,1975,1976],{},"同一顆模型、同樣的位元數，只差一張校準表，從 0 題變成三位數。"," 這就是「量化不只是壓縮，是配方」的意思。",[131,1979,1981],{"id":1980},"另一種壞法少了-chat-template","另一種壞法：少了 chat template",[111,1983,1984],{},"七個 suspects 之外，作者還測了 22 個 Hugging Face 社群檔案，其中兩個（TheBloke 的 deepseek-coder-6.7B Q3_K_M／Q4_K_M）在普查時只對 5、6 題。",[111,1986,1987,1988,1991],{},"原因不是量化，是",[413,1989,1990],{},"檔案沒有附 chat template","。Ollama 找不到 template，就用預設的 ChatML 去套，跟模型原本的指令格式不合——30 個回答裡有 28 個自己接出幻覺的對話輪。",[111,1993,1994,1995,1998],{},"補上正確 template 之後：",[413,1996,1997],{},"98 與 101 題","（滿分 164）。",[111,2000,2001],{},"所以「跑得起來」跟「它能好好回答」之間，還隔著一個 template。",[131,2003,2005],{"id":2004},"我們今天查了這些-tag-還在","我們今天查了：這些 tag 還在",[111,2007,2008],{},"文章不能只講九月的事。我們拿稽核資料裡的 blob digest，去比對 Ollama registry 今天的 manifest：",[194,2010,2011,2026],{},[197,2012,2013],{},[200,2014,2015,2017,2020,2023],{},[203,2016,1919],{},[203,2018,2019],{},"稽核時的 digest",[203,2021,2022],{},"今天",[203,2024,2025],{},"狀態",[219,2027,2028,2042,2054,2066,2078,2090,2102],{},[200,2029,2030,2033,2036,2039],{},[224,2031,2032],{},"qwen2.5-coder:3b-instruct-q2_K",[224,2034,2035],{},"14c4247b8ce4…",[224,2037,2038],{},"相同",[224,2040,2041],{},"還在",[200,2043,2044,2047,2050,2052],{},[224,2045,2046],{},"qwen2.5-coder:3b-instruct-q3_K_S",[224,2048,2049],{},"81d167af7fa3…",[224,2051,2038],{},[224,2053,2041],{},[200,2055,2056,2059,2062,2064],{},[224,2057,2058],{},"qwen2.5-coder:3b-instruct-q3_K_M",[224,2060,2061],{},"9bb5d28b842a…",[224,2063,2038],{},[224,2065,2041],{},[200,2067,2068,2071,2074,2076],{},[224,2069,2070],{},"qwen2.5-coder:3b-instruct-q3_K_L",[224,2072,2073],{},"7a88dc11faae…",[224,2075,2038],{},[224,2077,2041],{},[200,2079,2080,2083,2086,2088],{},[224,2081,2082],{},"phi3.5:3.8b-mini-instruct-q2_K",[224,2084,2085],{},"54d47caa8bf3…",[224,2087,2038],{},[224,2089,2041],{},[200,2091,2092,2095,2098,2100],{},[224,2093,2094],{},"llama3.2:1b-instruct-q2_K",[224,2096,2097],{},"d1ede02a2111…",[224,2099,2038],{},[224,2101,2041],{},[200,2103,2104,2107,2110,2112],{},[224,2105,2106],{},"llama3.2:1b-instruct-q3_K_S",[224,2108,2109],{},"81881dcc12ec…",[224,2111,2038],{},[224,2113,2041],{},[111,2115,2116,2119,2120,2123],{},[413,2117,2118],{},"七個都還在，digest 跟當時一模一樣。"," 這不是歷史事件，是你今天 ",[143,2121,2122],{},"ollama pull"," 就可能拿到的檔案。",[111,2125,2126,2127,2130,2131,2134],{},"（查法：",[143,2128,2129],{},"https:\u002F\u002Fregistry.ollama.ai\u002Fv2\u002Flibrary\u002F\u003Cmodel>\u002Fmanifests\u002F\u003Ctag>","，比對 ",[143,2132,2133],{},"application\u002Fvnd.ollama.image.model"," 那層的 digest。腳本在文末。）",[131,2136,2137],{"id":2137},"你現在該做的三件事",[111,2139,2140,2143,2144,2146,2147,2150],{},[413,2141,2142],{},"一、篩選。"," 作者釋出了 ",[143,2145,1819],{},"：對每個 tag 跑 15 題螢幕測試。成本大約 ",[413,2148,2149],{},"30 秒生成","（4090、不含下載）。66 個通過螢幕測試的重測，沒有一個是壞的——所以它是個堪用的預檢。",[111,2152,2153,2156,2157,2160],{},[413,2154,2155],{},"二、如果你自己發佈量化模型","：補上 imatrix，然後",[413,2158,2159],{},"在發佈前跑一次功能測試","。registry 不會幫你測。",[111,2162,2163,2166,2167,2170,2171,2174],{},[413,2164,2165],{},"三、檢查 chat template。"," ",[143,2168,2169],{},"ollama show \u003Cmodel> --template","；如果印出 ",[143,2172,2173],{},"{{ .Prompt }}","，代表沒有 template，模型可能用錯格式回答。",[131,2176,875],{"id":875},[835,2178,2179,2182,2185,2192],{},[838,2180,2181],{},"2.3% 是「Ollama 官方庫、能寫程式、8GB 以下」這個子集的比例，不是所有量化模型",[838,2183,2184],{},"只測程式碼能力（HumanEval+）；perplexity 只能當預檢，不能當功能證明",[838,2186,2187,2188,2191],{},"七個是",[413,2189,2190],{},"下限","：deepseek-coder-1.3B 全系列（含 fp16）都只對 1–6 題，無法用群組基線判定，作者沒有把它們算進去",[838,2193,2194],{},"作者是獨立研究者，資料與工具都公開，但這是單一 registry、單一 runtime 版本的結果",[131,2196,2197],{"id":2197},"結論",[111,2199,2200],{},[413,2201,2202],{},"可重現，不等於可用。",[111,2204,2205,2206,2208],{},"檔案一致、checksum 通過、",[143,2207,1868],{}," 正常——這些都只證明「檔案是你以為的那個檔案」，不證明「模型會做事」。",[111,2210,2211,2212,2215],{},"所以問題不是「量化能不能用」，是它",[413,2213,2214],{},"有沒有被測過","。目前答案是：registries 沒有，所以你得自己測。",[2217,2218],"hr",{},[111,2220,2221],{},[2222,2223,2224,2225,2230,2231,2230,2236,872],"em",{},"原始資料：",[118,2226,2229],{"href":2227,"rel":2228},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.05881",[122],"arXiv 2609.05881v3","、",[118,2232,2235],{"href":2233,"rel":2234},"https:\u002F\u002Fgithub.com\u002Faditi-p31\u002Fquantcheck",[122],"quantcheck（稽核資料與工具）",[118,2237,2240],{"href":2238,"rel":2239},"https:\u002F\u002Fgithub.com\u002Follama\u002Follama\u002Fissues\u002F18252",[122],"Ollama issue #18252",[111,2242,2243],{},[2222,2244,2245,2246,2249,2250,2253],{},"我們比對 digest 的腳本：",[143,2247,2248],{},"scripts\u002Fcheck-ollama-tags.ts","。跑 ",[143,2251,2252],{},"npm run check:ollama -- --list"," 會列出這七個 tag 今天的 digest 與稽核值是否一致。",{"title":730,"searchDepth":755,"depth":755,"links":2255},[2256,2257,2258,2259,2260,2261,2262,2263,2264],{"id":133,"depth":755,"text":133},{"id":1845,"depth":755,"text":1846},{"id":1876,"depth":755,"text":1876},{"id":1906,"depth":755,"text":1907},{"id":1980,"depth":755,"text":1981},{"id":2004,"depth":755,"text":2005},{"id":2137,"depth":755,"text":2137},{"id":875,"depth":755,"text":875},{"id":2197,"depth":755,"text":2197},"\u002Fimages\u002Fblog\u002Fbroken-on-arrival-quantized-models.svg","量化模型檔案完整但功能失效的示意圖，checksum 打勾、程式碼全錯","2026-11-02","有人把 Ollama 官方庫 305 個程式碼模型全跑了一遍，發現 7 個低 bit 量化版幾乎全錯。但檔案沒有壞——重建回來 tensor 完全一致，所以 checksum 抓不到。這篇講它怎麼發生的、你現在能做的三件事。",{},"\u002Fimages\u002Fblog\u002Fbroken-on-arrival-quantized-models.jpg","\u002Fblog\u002Fzh-tw\u002Fbroken-on-arrival-quantized-models",{"title":1809,"description":2268},"broken-on-arrival-quantized-models","blog\u002Fzh-TW\u002Fbroken-on-arrival-quantized-models",[1007,1008,1009],"iM5l1m1nXRGY-pGG2MmAYggyxY9V2uAe132PQbEixw8",{"id":2278,"title":2279,"author":103,"body":2280,"cover":3077,"coverAlt":3078,"date":3079,"description":3080,"draft":997,"extension":998,"featured":997,"lang":999,"meta":3081,"navigation":403,"ogImage":3082,"path":3083,"seo":3084,"slug":3085,"stem":3086,"tags":3087,"translationKey":3085,"updated":1010,"__hash__":3091},"blog\u002Fblog\u002Fzh-TW\u002Fweight-edit-loop-repair.md","它沒有變聰明，只是不再打轉",{"type":105,"value":2281,"toc":3066},[2282,2287,2289,2292,2298,2307,2310,2325,2328,2331,2422,2425,2428,2431,2434,2437,2443,2446,2450,2453,2456,2459,2469,2472,2489,2492,2495,2503,2506,2513,2524,2528,2531,2534,2581,2592,2598,2601,2604,2607,2610,2614,2617,2620,2623,2626,2629,2634,2640,2643,2651,2654,2658,2661,2670,2673,2676,2815,2818,2821,2824,2827,2830,2835,2841,2844,2847,2850,2853,2856,2859,2868,2871,2874,2912,2918,2921,2932,2935,2939,2942,2953,2956,2959,2966,2971,2978,2984,2987,2995,2998],[108,2283,2284],{},[111,2285,2286],{},"這篇的數字來自原文（部落格貼文，不是同儕審查論文）與公開 logs、Qwen 的 model card、Hugging Face 上的重建版重跑，以及硅星人Pro 按 logs 做的逐題重算。逐題配對那一段是二手重算，我會標出來。我們沒有在自己的機器上重跑那 50 次前向。",[131,2288,133],{"id":133},[111,2290,2291],{},"你在筆電上跑一顆 0.8B 的小模型，給它一題國中數學。它想了三分鐘，輸出到上限被切斷；回頭看，整段在重複同兩句話。",[111,2293,2294,2295],{},"你的結論大概是：模型太小。這篇要給你的是一個更準的分界——",[413,2296,2297],{},"分數變好，是模型變聰明，還是它終於講得完？",[111,2299,2300,2301,2306],{},"十一月二日，四位 MIT CSAIL 的研究者公開",[118,2302,2305],{"href":2303,"rel":2304},"https:\u002F\u002Frikiyat.github.io\u002Fblog\u002Fweight-space-programming\u002F",[122],"一組實驗","。他們讓 GPT-6 Astra 直接修改 Qwen3.5-0.8B 的權重，只動了 0.016%。",[111,2308,2309],{},"同一顆模型，MATH-500 從 7.9% 變成 48.1%。",[111,2311,2312,2313,2318,2319,2324],{},"那則貼文在 ",[118,2314,2317],{"href":2315,"rel":2316},"https:\u002F\u002Fnews.ycombinator.com\u002Fitem?id=50341174",[122],"Hacker News"," 拿到將近 3,000 分、900 則留言。中文媒體的標題則是「爆改 Qwen 大腦」這一路（",[118,2320,2323],{"href":2321,"rel":2322},"https:\u002F\u002F36kr.com\u002Fp\u002F4056494372181124",[122],"例","）。",[111,2326,2327],{},"拆開來看，那個六倍買到的不是數學能力，是「把話講完」。而這兩件事在 benchmark 表上長得一樣，除非你多記一欄。",[131,2329,2330],{"id":2330},"先讀最後一欄",[194,2332,2333,2352],{},[197,2334,2335],{},[200,2336,2337,2340,2343,2346,2349],{},[203,2338,2339],{},"Qwen3.5-0.8B",[203,2341,2342],{},"MATH-500",[203,2344,2345],{},"GSM8K",[203,2347,2348],{},"IFEval",[203,2350,2351],{},"撞 32,768 token 上限",[219,2353,2354,2371,2388,2405],{},[200,2355,2356,2359,2362,2365,2368],{},[224,2357,2358],{},"原始",[224,2360,2361],{},"7.9%",[224,2363,2364],{},"13.8%",[224,2366,2367],{},"21.0%",[224,2369,2370],{},"93.0%",[200,2372,2373,2376,2379,2382,2385],{},[224,2374,2375],{},"改，10 次前向",[224,2377,2378],{},"12.3%",[224,2380,2381],{},"25.7%",[224,2383,2384],{},"49.0%",[224,2386,2387],{},"87.2%",[200,2389,2390,2393,2396,2399,2402],{},[224,2391,2392],{},"改，25 次前向",[224,2394,2395],{},"14.1%",[224,2397,2398],{},"25.1%",[224,2400,2401],{},"39.9%",[224,2403,2404],{},"86.4%",[200,2406,2407,2410,2413,2416,2419],{},[224,2408,2409],{},"改，50 次前向",[224,2411,2412],{},"48.1%",[224,2414,2415],{},"67.1%",[224,2417,2418],{},"62.3%",[224,2420,2421],{},"33.1%",[111,2423,2424],{},"前向（forward pass）就是讓模型真的回答一批題目，看它輸出什麼。不是訓練，也不給梯度。",[111,2426,2427],{},"四列用的是同一個方法，但三次交出的是三份不同的補丁，不是同一招多給預算。差別在於 Astra 被允許跑幾次前向。",[111,2429,2430],{},"10 次：12.3。25 次：14.1。50 次：48.1。",[111,2432,2433],{},"看前三列，你會說「預算不夠」。但最後一欄講的是另一件事。",[111,2435,2436],{},"10 次和 25 次時，撞上上限的數學題還有 87% 和 86%。到 50 次，掉到 33.1%。",[111,2438,2439,2442],{},[413,2440,2441],{},"分數是跟著截斷率走的。"," 前兩次沒有把迴圈修掉，分數就留在原地；把撞上限的比例壓下來，分數才動。",[111,2444,2445],{},"有一條線不聽這個解釋：IFEval 在 10 次前向就從 21.0 跳到 49.0，那時截斷率還有 87.2%。IFEval 的答案短，截斷對它的影響小。單次執行的結果，我不會拿它下結論。",[131,2447,2449],{"id":2448},"它改的是什麼兩層一個方向十三萬個權重","它改的是什麼：兩層、一個方向、十三萬個權重",[111,2451,2452],{},"Astra 做的事可以拆成三步。",[111,2454,2455],{},"第一步，看 96 道練習題的回答。其中 85 道撞到 32,768 上限，撞上去的內容幾乎都在重複。",[111,2457,2458],{},"第二步，把這個重複追到第 15 層與第 19 層的 MLP。MLP 是每個 transformer 區塊裡的前饋層，這兩層的輸出一直在同一個方向上推。",[111,2460,2461,2462,2465,2466,2468],{},"第三步，從這兩層的 ",[143,2463,2464],{},"down_proj"," 各挑 64 個輸出欄位，把那個方向減掉。",[143,2467,2464],{}," 是前饋層最後把結果投影回模型寬度的那個矩陣。",[111,2470,2471],{},"2 × 64 × 1,024 = 131,072 個權重，約占 8 億的 0.016%。",[111,2473,2474,2475,2480,2481,2484,2485,2488],{},"Hugging Face 上有人照公開的 diff ",[118,2476,2479],{"href":2477,"rel":2478},"https:\u002F\u002Fhuggingface.co\u002FRikiyaT\u002FQwen3.5-0.8B-weight-space-programming",[122],"重建了檢查點","。公開 logs 裡的腳本逐個 tensor 對 sha256：只有 ",[143,2482,2483],{},"layers.15.mlp.down_proj"," 與 ",[143,2486,2487],{},"layers.19.mlp.down_proj"," 兩個張量不同。",[111,2490,2491],{},"我認為這是最乾淨的一種驗證。不是「他們說改了兩個張量」，是「重建出來的檔案只有兩個張量不同」。",[111,2493,2494],{},"比喻的話：模型不是不會算，它是在某個地方一直被推回原點，繞著同樣兩三句話轉。那個「推回去的力」，在權重裡是一個方向。減掉它，模型就能把已經想到的答案寫完（圖 1）。",[111,2496,2497,2498,2502],{},"Astra 還自己做了一個對照。取一個與目標方向正交、改動量相同的隨機方向減掉，成績只動 1.5 個百分點（",[118,2499,2501],{"href":2321,"rel":2500},[122],"覆盤","）。所以動的是方向，不是「隨便改一點權重都會變好」。",[111,2504,2505],{},"這不是全新的技術路線。把「拒絕回答」定位到單一方向的研究、ROME 與 MEMIT 那種「找到一組權重、改掉一件事」的做法，都在同一個家族裡。",[111,2507,2508,2509,2512],{},"新的部分是",[413,2510,2511],{},"誰來找那個方向。"," 這次是一個通用模型自己找，而且會把理由寫出來。",[393,2514,397,2516,397,2521],{"className":2515},[396],[399,2517],{"src":2518,"poster":2519,"autoPlay":403,"loop":403,"muted":403,"playsInline":403,"ariaLabel":2520},"\u002Fimages\u002Fblog\u002Fweight-edit-loop-repair\u002Floop-repair.mp4","\u002Fimages\u002Fblog\u002Fweight-edit-loop-repair\u002Floop-repair-poster.jpg","左邊是原始模型反覆輸出、撞上 32,768 上限；右邊是減掉一個方向之後，答案在預算內寫完",[406,2522,2523],{},"圖 1：同一顆模型。左邊一直被推回原點，93% 的數學題撞上上限；右邊減掉兩層 MLP 裡的方向，撞上限的比例掉到 33%。",[131,2525,2527],{"id":2526},"_579-題裡576-題原本只是講不完","579 題裡，576 題原本只是講不完",[111,2529,2530],{},"這是最關鍵、也最容易被誤讀的一段。",[111,2532,2533],{},"硅星人Pro 把公開 logs 的逐題紀錄重算一次。他們挑的 1,152 道題，是練習題裡保留起來、沒有拿去挑補丁的那批。",[194,2535,2536,2552],{},[197,2537,2538],{},[200,2539,2540,2543,2546,2549],{},[203,2541,2542],{},"原本的狀態",[203,2544,2545],{},"題數",[203,2547,2548],{},"改之前答對",[203,2550,2551],{},"改之後答對",[219,2553,2554,2567],{},[200,2555,2556,2559,2562,2565],{},[224,2557,2558],{},"有正常結束",[224,2560,2561],{},"114",[224,2563,2564],{},"93.9%",[224,2566,2564],{},[200,2568,2569,2572,2575,2578],{},[224,2570,2571],{},"被上限截斷",[224,2573,2574],{},"1,038",[224,2576,2577],{},"1.4%",[224,2579,2580],{},"56.7%",[835,2582,2583,2586,2589],{},[838,2584,2585],{},"從錯變對 579 題，其中 576 題原本是被截斷的",[838,2587,2588],{},"從對變錯 5 題",[838,2590,2591],{},"平均回答長度從 30,868 token 降到 22,037",[111,2593,2594,2597],{},[413,2595,2596],{},"已經答完的 114 題，正確率一位小數都沒動。"," 114 題裡 3 題從錯變對、3 題從對變錯，剛好打平。",[111,2599,2600],{},"579 個救援裡，576 個來自那 1,038 道本來沒講完的。真正的進步在那裡。",[111,2602,2603],{},"這件事不是誰抓到破綻。原文的表格最後一欄就是 93% 到 33.1%，Astra 自己的最終報告也寫了 576／579。",[111,2605,2606],{},"作者沒有隱瞞：logs、檢查點、失敗的實驗都公開了。他們只是沒有把「已經答完的題目有沒有變」算成一行。",[111,2608,2609],{},"補一句：這張表是依 logs 重算出來的，不是原文的表格。我們沒有自己重算。",[131,2611,2613],{"id":2612},"不會打轉的模型它修不動","不會打轉的模型，它修不動",[111,2615,2616],{},"同樣的方法，換到 Gemma 3 270M 就失效了。",[111,2618,2619],{},"這顆模型很少打轉：只有 8.5% 的回答撞到上限。Qwen 那顆是 93%。",[111,2621,2622],{},"Astra 讀它的失敗，找到的是普通的算錯與代數步驟錯。沒有一個可以關掉的零件。",[111,2624,2625],{},"改完之後，MATH-500 落在 7.0% 到 7.7% 之間，改之前是 6.8%。Astra 自己的報告把這個幅度標成「未達確立」。",[111,2627,2628],{},"原文那句話，是整篇最重要的：",[108,2630,2631],{},[111,2632,2633],{},"Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked.",[111,2635,2636,2637],{},"翻成白話：",[413,2638,2639],{},"改權重可以移除擋住能力的障礙，但不能加進模型本來沒有的能力。",[111,2641,2642],{},"這句話對你有直接用途。你的模型表現差，先問是哪一種差。",[835,2644,2645,2648],{},[838,2646,2647],{},"它會算，但講不完（一直重複、撞上限）→ 移除障礙有效",[838,2649,2650],{},"它算錯 → 移除障礙無效，換模型或換任務",[111,2652,2653],{},"第一種，可以先從設定下手，不必換模型。第二種，別在權重上花時間。",[131,2655,2657],{"id":2656},"被批的不是結果是包裝","被批的不是結果，是包裝",[111,2659,2660],{},"這篇被罵得很兇。但先把話講清楚：被批的是包裝，不是結果。",[111,2662,2663,2664,2669],{},"spaCy 的作者 Matthew Honnibal 寫了一篇",[118,2665,2668],{"href":2666,"rel":2667},"https:\u002F\u002Fhonnibal.dev\u002Fblog\u002Fastra-read-write-misleading",[122],"質疑","，用魔術師比喻。",[111,2671,2672],{},"同一場戲，第一次用特殊道具，第二次改用藏東西，第三次用強迫選擇。觀眾以為只有一種手法，於是每看一次都像推翻了上一個假設。",[111,2674,2675],{},"他的證據是一張表，把八個實驗按「Astra 實際拿到什麼」排開：",[194,2677,2678,2697],{},[197,2679,2680],{},[200,2681,2682,2685,2688,2691,2694],{},[203,2683,2684],{},"實驗",[203,2686,2687],{},"對象",[203,2689,2690],{},"只看原始權重",[203,2692,2693],{},"用 Python",[203,2695,2696],{},"跑過模型",[219,2698,2699,2714,2728,2742,2756,2770,2785,2800],{},[200,2700,2701,2704,2707,2710,2712],{},[224,2702,2703],{},"1 猜遊戲",[224,2705,2706],{},"1.7 萬",[224,2708,2709],{},"✓",[224,2711,553],{},[224,2713,553],{},[200,2715,2716,2719,2722,2724,2726],{},[224,2717,2718],{},"2 猜形狀",[224,2720,2721],{},"2.4 萬",[224,2723,2709],{},[224,2725,553],{},[224,2727,553],{},[200,2729,2730,2733,2736,2738,2740],{},[224,2731,2732],{},"3 猜語言與複製頭",[224,2734,2735],{},"716 萬",[224,2737,553],{},[224,2739,2709],{},[224,2741,553],{},[200,2743,2744,2747,2750,2752,2754],{},[224,2745,2746],{},"4 修卡住的車",[224,2748,2749],{},"17 萬",[224,2751,2709],{},[224,2753,553],{},[224,2755,553],{},[200,2757,2758,2761,2764,2766,2768],{},[224,2759,2760],{},"5 從零寫 MinAtar",[224,2762,2763],{},"2 萬",[224,2765,553],{},[224,2767,553],{},[224,2769,553],{},[200,2771,2772,2775,2778,2780,2782],{},[224,2773,2774],{},"6 從零寫圍棋",[224,2776,2777],{},"148 萬",[224,2779,553],{},[224,2781,2709],{},[224,2783,2784],{},"✓（0–50 次）",[200,2786,2787,2790,2793,2795,2797],{},[224,2788,2789],{},"7 找出阿諛模型",[224,2791,2792],{},"2.7 億",[224,2794,553],{},[224,2796,2709],{},[224,2798,2799],{},"✓（只看 attention）",[200,2801,2802,2805,2808,2810,2812],{},[224,2803,2804],{},"8 修 Qwen 數學",[224,2806,2807],{},"約 8 億",[224,2809,553],{},[224,2811,2709],{},[224,2813,2814],{},"✓（50 次前向）",[111,2816,2817],{},"（表依 Honnibal 的整理改寫，數字用他的級距。實驗 4 另外拿到一段失敗錄影。實驗 5 的權重從 0 開始，沒有東西可讀。實驗 7 只看權重時四次全錯，加上 attention 工具才三對三。）",[111,2819,2820],{},"真正「只看原始權重、不給工具、不跑模型」的，只有實驗 1、2、4。對象是 1.7 萬到 17 萬個權重的小網路。",[111,2822,2823],{},"所有碰到真語言模型的實驗都用 Python。Qwen 那組還被允許跑 50 次前向。",[111,2825,2826],{},"摘要卻把三種能力寫成一句話：「只看原始權重，就能理解並修改另一個模型。」讀者自然會以為是同一招。",[111,2828,2829],{},"還有一件事更硬。那次數學實驗的提示詞裡，作者直接寫了：",[108,2831,2832],{},[111,2833,2834],{},"if the model often fails because it repeats itself and never finishes, finding and fixing what drives that repetition counts",[111,2836,2837,2838],{},"也就是說，",[413,2839,2840],{},"「重複」被當成有效修改的例子，寫在任務說明裡。",[111,2842,2843],{},"他們也做了一個拿掉這個提示的版本（v4）。10 次前向時不再交出修改，25 次只到 14.1%，而且沒有跑 50 次的版本。",[111,2845,2846],{},"所以「沒有提示時能不能到 48.1」，公開資料回答不了。",[111,2848,2849],{},"（v4 除了拿掉失敗範例，還改了練習題與驗證題的分法和流程。這個對照不乾淨，但方向值得記住。）",[111,2851,2852],{},"也該給作者記一筆。貼文公開後不久，他們改了摘要：把「遠超過人類或傳統演算法」拿掉，換成比較保留的說法。",[111,2854,2855],{},"負面實驗照登，logs 全部公開，檢查點也放出來。",[111,2857,2858],{},"我的判斷：能力是真的。代價與包裝被放大，但證據鏈沒有斷。",[111,2860,2861,2862,2867],{},"Hugging Face 上的",[118,2863,2866],{"href":2864,"rel":2865},"https:\u002F\u002Fhuggingface.co\u002FprithivMLmods\u002FQwen3.5-0.8B-weight-space-programming-GGUF",[122],"重建版","重跑 500 題，編輯後拿到 51.0%，撞上限 31.6%。換一個人、換一張卡，結論沒變。",[131,2869,2870],{"id":2870},"你明天可以做的四件事",[111,2872,2873],{},"不用等下一篇論文。任何一次「改完之後分數變好」，都可以用這四步驗。",[944,2875,2876,2889,2895,2906],{},[838,2877,2878,2884,2885,2888],{},[413,2879,2880,2881,872],{},"加一欄 ",[143,2882,2883],{},"finish_reason"," 那是 API 回傳的結束原因，等於 ",[143,2886,2887],{},"length"," 就是被上限截斷。這個比例超過 10%，就不要直接比總分。",[838,2890,2891,2894],{},[413,2892,2893],{},"看截斷回應的尾巴。"," 8-gram 重複率，就是連續八個 token 一模一樣的比例。太高，表示它在繞圈，不是題目太難。",[838,2896,2897,2900,2901,2905],{},[413,2898,2899],{},"核設定。"," 思考模式開關、temperature、token 上限，跟 model card 對一次。這篇的基準是強制開思考、temperature 0.6、presence penalty 1.5、上限 32,768；",[118,2902,2339],{"href":2903,"rel":2904},"https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3.5-0.8B",[122]," 的預設是非思考模式。",[838,2907,2908,2911],{},[413,2909,2910],{},"加一個不會打轉的對照模型。"," 它沒進步，你看到的就不是能力提升。",[111,2913,2914,2915,872],{},"四個步驟都不需要幾十小時的模型對話，只需要一個 ",[143,2916,2917],{},"if",[111,2919,2920],{},"那什麼時候真的值得改權重？三種情況。",[835,2922,2923,2926,2929],{},[838,2924,2925],{},"病理已經被寫進 model card，像這顆模型的迴圈，Qwen 自己警告過",[838,2927,2928],{},"你沒有資料或算力做微調",[838,2930,2931],{},"你接受它現在是研究階段：一次成功的編輯花了 80 次程式執行、158 輪對話，接近一整天",[111,2933,2934],{},"作者自己也說，還沒證明比微調便宜。",[131,2936,2938],{"id":2937},"三個問題加上一句提醒","三個問題，加上一句提醒",[111,2940,2941],{},"下次看到「某個方法讓模型變強六倍」，先問三題。",[835,2943,2944,2947,2950],{},[838,2945,2946],{},"分母是什麼？（基準是在哪個設定下量的）",[838,2948,2949],{},"上限是多少？（有多少題根本沒講完）",[838,2951,2952],{},"有沒有對照組？（換一顆不會打轉的模型，還成立嗎）",[111,2954,2955],{},"這三題擋不掉所有誤讀，但能擋掉大部分新聞。",[111,2957,2958],{},"如果你正在選地端模型，或評估一個「調完之後變很好」的內部結果，我們做過類似的事。",[111,2960,2961,2962,2324],{},"把能跑、能回答、能通過驗收拆成四層（",[118,2963,2965],{"href":2964},"\u002Fblog\u002Frunning-is-not-done","「跑得起來」跟「真的幹完活」是四件事",[111,2967,2968,2969,2324],{},"同一個道理換到量化模型，是「檔案本身可能是壞的」（",[118,2970,955],{"href":954},[111,2972,2973,2974,2324],{},"而「把移除障礙寫成能力進化」這個模式，我們也寫過一次（",[118,2975,2977],{"href":2976},"\u002Fblog\u002Feth-refuted-not-p-np","「推翻 ETH」不是「攻破 P=NP」",[111,2979,2980,2981],{},"這三篇講的是同一件事：",[413,2982,2983],{},"分數變好，跟模型變強，是兩個不同的句子。",[111,2985,2986],{},"要把它們分開，只需要一欄資料。而那一欄，你今天的 pipeline 其實就有了。",[111,2988,2989,2990,2994],{},"想把這件事變成團隊的例行檢查？我們做",[118,2991,2993],{"href":2992},"\u002Fservices","地端 AI 導入","時，第一件事就是這個：先量，再決定要不要換模型。",[131,2996,2997],{"id":2997},"資料來源",[835,2999,3000,3008,3017,3029,3037,3046,3059],{},[838,3001,3002,3003,3007],{},"原始貼文：",[118,3004,3006],{"href":2303,"rel":3005},[122],"Sparks of Weight-Space Programming","（Rikiya Takehi、Jayden Teoh、Alex Zhang、Omar Khattab，2026-11-02；未同儕審查）",[838,3009,3010,3011,3016],{},"完整 logs：",[118,3012,3015],{"href":3013,"rel":3014},"https:\u002F\u002Fgithub.com\u002FRikiyaT\u002Flm_understand-logs",[122],"RikiyaT\u002Flm_understand-logs","，1,125 次模型調用",[838,3018,3019,3020,3024,3025],{},"檢查點：",[118,3021,3023],{"href":2477,"rel":3022},[122],"RikiyaT\u002FQwen3.5-0.8B-weight-space-programming","；重建版重跑數據見 ",[118,3026,3028],{"href":2864,"rel":3027},[122],"GGUF 卡",[838,3030,3031,3032,3036],{},"逐題重算（1,152 題配對、579／576、隨機方向對照）：硅星人Pro〈",[118,3033,3035],{"href":2321,"rel":3034},[122],"刷屏的「Astra 改造 Qwen 大腦」跟我們想要的 RSI 沒多大關係","〉，2026-11-05",[838,3038,3039,3040,3045],{},"另一份獨立整理：",[118,3041,3044],{"href":3042,"rel":3043},"https:\u002F\u002Ftraictory.com\u002Fnews\u002F2026-11-04-mit-weight-space-programming",[122],"traictory","，2026-11-04",[838,3047,3048,3049,3053,3054],{},"批評：",[118,3050,3052],{"href":2666,"rel":3051},[122],"Honnibal","，2026-11-03；",[118,3055,3058],{"href":3056,"rel":3057},"https:\u002F\u002Fsimonwillison.net\u002F2026\u002FNov\u002F3\u002Fqwen-can-interpret-model-weights\u002F",[122],"Simon Willison 的整理",[838,3060,3061,3062],{},"採樣設定與預設模式：",[118,3063,3065],{"href":2903,"rel":3064},[122],"Qwen3.5-0.8B model card",{"title":730,"searchDepth":755,"depth":755,"links":3067},[3068,3069,3070,3071,3072,3073,3074,3075,3076],{"id":133,"depth":755,"text":133},{"id":2330,"depth":755,"text":2330},{"id":2448,"depth":755,"text":2449},{"id":2526,"depth":755,"text":2527},{"id":2612,"depth":755,"text":2613},{"id":2656,"depth":755,"text":2657},{"id":2870,"depth":755,"text":2870},{"id":2937,"depth":755,"text":2938},{"id":2997,"depth":755,"text":2997},"\u002Fimages\u002Fblog\u002Fweight-edit-loop-repair.svg","標題「它沒有變聰明，只是不再打轉」的技術筆記封面","2026-11-08","一個 0.8B 小模型改掉 0.016% 的權重，MATH-500 從 7.9% 跳到 48.1%。但逐題比對後，救回來的 579 題裡有 576 題原本是被 32,768 token 上限截斷的；已經答完的 114 題，正確率完全沒變。這篇把六倍拆開，給你四步檢查——下次看到分數暴漲時，先看截斷率，再看對照組。",{},"\u002Fimages\u002Fblog\u002Fweight-edit-loop-repair.jpg","\u002Fblog\u002Fzh-tw\u002Fweight-edit-loop-repair",{"title":2279,"description":3080},"weight-edit-loop-repair","blog\u002Fzh-TW\u002Fweight-edit-loop-repair",[1008,3088,3089,3090],"reasoning-models","evaluation","interpretability","WA7IcJhsMEwChZwPlmimkW-vQze8P_jzZWU7jGg1Yxw",1795088284302]