[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-rrfp-readiness-driven-pipeline-training-zh":3,"article-related-rrfp-readiness-driven-pipeline-training-zh":36,"series-research-eda7a80a-b234-4ada-90d1-a37b144251dc":89},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":18,"translated_content":10,"views":19,"is_premium":20,"created_at":21,"updated_at":21,"cover_image":11,"published_at":22,"rewrite_status":23,"rewrite_error":10,"rewritten_from_id":24,"slug":25,"category":26,"related_article_id":27,"status":28,"google_indexed_at":29,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":30,"topic_cluster_id":34,"embedding":35,"is_canonical_seed":20},"eda7a80a-b234-4ada-90d1-a37b144251dc","RRFP 讓管線訓練跟著就緒跑","\u003Cp data-speakable=\"summary\">RRFP 把管線平行訓練改成先跑已就緒工作，減少 runtime 變動造成的閒置空泡。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>研究機構\u003C\u002Fstrong>：arXiv 摘要未明確標註\u003C\u002Fli>\u003Cli>\u003Cstrong>核心數據\u003C\u002Fstrong>：多模態工作負載最高 2.77× 加速\u003C\u002Fli>\u003Cli>\u003Cstrong>突破點\u003C\u002Fstrong>：以就緒度優先派工\u003C\u002Fli>\u003C\u002Ful>\u003Cp>管線平行訓練一直是大型模型擴展的重要手段，但它有一個老問題：排程寫得再漂亮，runtime 一旦有變動，整條管線還是可能卡住。這篇論文要解的，就是這種「紙上順、實際慢」的落差。\u003C\u002Fp>\u003Cp>RRFP，全名是 Runtime-Readiness-First Pipeline。它的核心不是重新發明訓練流程，而是改變 runtime 看待排程的方式。論文主張，當計算與通訊在執行時出現變動時，固定順序會讓 stage 等下一個指定工作，明明有別的工作已經能跑，卻還是閒著。\u003C\u002Fp>\u003Cp>對開發者來說，這不是抽象的系統細節。它會直接\u003Ca href=\"\u002Fnews\u002Fwembanyama-stat-page-turns-into-recap-zh\">變成\u003C\u002Fa> idle bubbles、\u003Ca href=\"\u002Ftag\u002Fgpu\">GPU\u003C\u002Fa> 利用率下降，還有整體訓練時間拉長。RRFP 想做的，就是把這些浪費吃回來。\u003C\u002Fp>\u003Ch2>這篇在修哪個痛點\u003C\u002Fh2>\u003Cp>論文指出，現有 pipeline 系統通常假設：前面規劃好的順序，就是 runtime 應該照做的順序。這個假設只有在任務就緒狀態和排程順序完全一致時才成立。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779172442474-n21q.png\" alt=\"RRFP 讓管線訓練跟著就緒跑\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>但實際訓練不是這樣。計算時間會飄，通訊也會飄。當下一個排程項目還沒準備好，stage 就只能等；可是一旁如果已經有別的任務 ready，固定順序也不會自動去接它。於是，管線看起來有條有理，實際上卻在空轉。\u003C\u002Fp>\u003Cp>RRFP 的出發點，就是把這個錯位拆掉。它不是讓 schedule 決定一切，而是讓「現在誰已經 ready」來決定先做誰。\u003C\u002Fp>\u003Ch2>RRFP 怎麼運作\u003C\u002Fh2>\u003Cp>這篇論文最重要的設計，是把 schedule 從「硬性命令」改成「提示」。也就是說，排程順序還在，但不再是 runtime 必須死守的唯一順序。runtime 會先看目前有哪些工作已經可以執行，再用 hint order 去排序這些 ready 的工作。\u003C\u002Fp>\u003Cp>這個改法看起來只是換了一個優先級，實際上卻改了控制模型。stage 不必因為下一個指定項目還沒好就停住，只要有其他可執行任務，就能先把硬體用起來。論文要解決的，是讓 pipeline 跟著「可做\u003Ca href=\"\u002Fnews\u002Fwhy-wembanyama-game-3-should-change-spurs-expectations-zh\">什麼\u003C\u002Fa>」走，而不是只跟著「原本打算先做什麼」走。\u003C\u002Fp>\u003Cp>為了支撐這種執行方式，RRFP 結合了三個機制：message-driven asynchronous communication、lightweight tensor-parallel coordination，以及 ready-set arbitration。摘要沒有把實作細節全部攤開，但方向很清楚：通訊要能異步、協調要夠輕、派工要能快速在 ready 集合裡做選擇。\u003C\u002Fp>\u003Cp>論文把 RRFP 做在 Megatron-based training framework 上。這代表它不是純理論排程，而是放進實際訓練框架裡驗證的 runtime 系統。\u003C\u002Fp>\u003Cp>另一個值得注意的點，是它並沒有把正確性當成可犧牲項目。摘要明確說，RRFP 仍然維持 training correctness。也就是說，它追求的是更聰明的執行順序，不是用近似結果換速度。\u003C\u002Fp>\u003Ch2>它實際證明了什麼\u003C\u002Fh2>\u003Cp>這篇摘要有提供數字，而且數字不小。評估涵蓋 language-only 和 multimodal workloads，規模最高到 128 GPUs。這很重要，因為 pipeline 的效率問題通常會在規模拉大後更明顯，小小的等待時間也會被放大成很可觀的浪費。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779172446424-cpbl.png\" alt=\"RRFP 讓管線訓練跟著就緒跑\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>結果方面，論文提到：在使用 BFW hint 的情況下，RRFP 在 language-only workloads 上最高可達 1.77× speedup，在 multimodal workloads 上最高可達 2.77× speedup。這種加速是來自 runtime 行為調整，不是改模型架構，也不是改訓練目標。\u003C\u002Fp>\u003Cp>論文也做了跨框架比較。摘要說，RRFP 搭配預設 BF hint，最多可以比「外部系統中最快的可用版本」快 1.84×，而且仍然保持訓練正確性。這句話的\u003Ca href=\"\u002Fnews\u002Fspurs-vs-timberwolves-game-5-takeaways-zh-tw-zh\">重點\u003C\u002Fa>是，它不是單純比誰更激進，而是把速度和 correctness 一起守住。\u003C\u002Fp>\u003Cp>但摘要也有明顯限制。它沒有公開完整的 benchmark 細節，沒有列出絕對 throughput、step time、記憶體成本，也沒有提供各個元件的 ablation 數字。換句話說，光看摘要可以知道方向和 headline 成果，但還不能完整判斷代價與 trade-off。\u003C\u002Fp>\u003Cp>另外，摘要也沒有把 BFW、BF 這些 hint 的差異講得很細。你可以知道 hint 會影響結果，但還看不出它們在不同工作負載下的敏感度有多高。\u003C\u002Fp>\u003Ch2>對開發者有什麼影響\u003C\u002Fh2>\u003Cp>如果你在做大型模型訓練，這篇論文傳達的訊號很直接：pipeline 排程不能只看理論順序，還要看 runtime 的就緒狀態。當工作負載有通訊延遲、計算波動，或兩者一起來時，固定排程很容易把可用算力浪費掉。\u003C\u002Fp>\u003Cp>這對混合工作負載特別有意義。論文把 language-only 和 multimodal 都納進來，表示它關注的不是單一模型類型，而是更廣義的訓練 runtime 問題。對系統工程師來說，這種 readiness-driven 的設計，可能比再微調一次靜態排程更有用。\u003C\u002Fp>\u003Cp>從框架設計角度看，RRFP 也提供了一個很實用的模式：把 schedule intent 和 dispatch order 拆開。前者保留規劃意圖，後者交給 runtime 根據 ready 狀態決定。這樣做的好處，是 runtime 能更靈活，但又不必變成一個很重的中央協調器。\u003C\u002Fp>\u003Cp>論文提到的 message-driven async communication 和 ready-set arbitration，就屬於這種低開銷的支撐機制。它們的目的不是增加複雜度，而是讓 runtime 有能力在不打亂正確性的前提下，把空等時間轉成實際工作。\u003C\u002Fp>\u003Ch2>限制與還沒回答的問題\u003C\u002Fh2>\u003Cp>摘要雖然給了漂亮的 speedup，但還留下不少空白。它沒有說 RRFP 在一般情況下會多多少開銷，也沒有說 hint order 的品質會不會影響結果。這些都會決定它能不能從研究原型走向更廣泛的系統實作。\u003C\u002Fp>\u003Cp>另外，摘要沒有說明收益到底更依賴 workload 類型、GPU 數量，還是 pipeline 形狀。雖然它同時測了 language-only 和 multimodal，也測到 128 GPUs，但這還不足以證明它對所有訓練堆疊都同樣有效。\u003C\u002Fp>\u003Cp>還有一個實務上的問題是：如果 readiness 資訊本身延遲或不夠準，RRFP 的優勢會不會被吃掉？摘要沒有回答。這代表它的效益很可能跟 runtime 觀測品質綁在一起。\u003C\u002Fp>\u003Cp>即便如此，方向仍然很清楚。隨著訓練工作負載越來越不穩定，死守固定 pipeline 順序會越來越像一種負擔。RRFP 想證明的是：只要 runtime 能根據就緒狀態動態派工，就有機會在不犧牲正確性的前提下，把硬體利用率拉回來。\u003C\u002Fp>\u003Ch2>結語\u003C\u002Fh2>\u003Cp>RRFP 把 pipeline-parallel training 的中心從「先排好」移到「先看誰 ready」。從摘要公開的數字來看，這不只是小修小補，而是能在多模態與大規模 GPU 環境下帶來明顯加速的 runtime 改造。\u003C\u002Fp>\u003Cp>對正在做大模型訓練的團隊來說，這篇論文的價值在於提醒一件事：訓練系統的競爭力，越來越不只在模型本身，也在 runtime 能不能跟上現實世界的變動。\u003C\u002Fp>","RRFP 把管線平行訓練改成先跑已就緒工作，減少 runtime 變動造成的閒置空泡，最高在多模態工作負載上快 2.77 倍。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.18750",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779172442474-n21q.png",[13,14,15,16,17],"pipeline parallelism","runtime scheduling","readiness-driven dispatch","Megatron","multimodal training","zh",2,false,"2026-05-19T06:33:31.287772+00:00","2026-05-19T06:33:31.266+00:00","done","853fe9db-d8ae-4eb5-b860-c08f7333e579","rrfp-readiness-driven-pipeline-training-zh","research","3440bae8-d711-472c-8861-ef8ea63d39e8","published","2026-05-19T09:00:33.107+00:00",[31,32,33],"把固定排程改成以就緒度優先，能減少 pipeline 閒置空泡。","摘要公開的結果顯示，最高可達 1.77× 到 2.77× 加速。","它維持訓練正確性，但摘要未公開完整 benchmark 與開銷細節。","0c35a120-52fc-41fc-afa3-d404eb934158","[0.00296292,0.020053986,0.035026286,-0.08881252,-0.027159635,-0.011495941,-0.0214171,0.0048894743,0.016228203,-0.008899702,0.0051576705,-0.011277806,0.03704715,0.01731654,0.10845327,0.0115155205,0.012820075,0.011734073,-0.0024651182,-0.0063549713,-0.010045282,0.008230375,-0.01513619,0.0027050148,-0.010709221,0.011800264,0.03685689,-0.0139580835,0.037808936,-0.000132037,-0.016207524,0.008231948,0.024524441,0.012702632,-0.013740967,0.0062255063,-0.0049950704,-0.028926224,0.016388103,0.0059949392,-0.025313117,-0.022638453,0.010280311,-0.0032544094,-0.025108233,-0.0023489771,-0.0029720971,-0.018660858,0.0027423971,0.025178984,-0.0035596346,0.0070768273,-0.02605745,-0.15693578,-0.005143978,-0.011246873,-0.0042021046,0.019504013,0.010568368,-0.005234828,-0.0019271475,0.012774942,-0.009172459,-0.016965048,-0.015046823,-0.015749164,0.02478067,-0.008156351,-0.0136348335,0.007148858,0.010639621,-0.010973255,-0.0005088685,-0.003932586,-0.004266791,-0.04435775,-0.007269761,0.04407595,-0.0069852923,0.026113663,0.019204343,-0.026364693,-0.021357076,-0.0070962375,0.008265743,-0.01251493,-0.0011075754,-0.012594249,0.005851073,0.0055551883,0.009227729,0.008180398,-0.0075651626,-0.017076707,0.0033860274,-0.028018998,0.025219547,0.0078078727,0.0013701677,-0.012245135,-0.009535128,-0.010960844,0.0035870592,0.020568604,-0.010871272,-0.010779296,-0.0038793054,0.006721289,0.0011491255,0.020046297,-0.0078426385,-0.034055967,0.022529792,0.022962775,0.0068657175,-0.123079106,-0.010481929,-0.010689147,-0.023317311,0.00078487676,-0.004041164,-0.009303076,0.017502,0.017671585,0.031181622,-0.006645977,0.006403019,-0.004224149,-0.0012546234,0.0034153564,-0.0078374185,0.003844782,0.00902993,0.023735223,0.013626422,0.017357899,0.012542389,-0.01318602,-0.014908778,-0.0056365235,-0.0059066517,0.0184443,-0.021759486,-0.0015037099,-0.014114494,-0.0044457023,-0.032317165,-0.009722711,0.01768423,-0.016824914,0.016371235,-0.021130696,0.0049481275,-0.005468165,-0.00077946024,-0.015544471,0.0052654543,-0.012144713,-0.029785391,0.008269752,-0.002886643,0.003941842,-0.0017225153,0.0038816403,-0.0048786015,0.025852067,-0.0027744982,0.012256735,0.0070977574,0.013995732,0.0081432825,-0.019635778,-0.024977,-0.0021591592,-0.00025850371,0.0077424888,-0.026103972,-0.0028912965,0.03492701,-0.019128626,0.009013915,-0.0028273242,-0.00041896972,-0.0066482443,0.0021441479,0.008017097,0.007983048,0.0091387145,0.009664832,-0.010008753,-0.033534557,0.0074157333,-0.011853355,-0.0038732707,0.0024179742,-0.038607292,0.018449223,-0.013513745,0.008664007,0.007127669,0.0012107395,-0.021237481,0.020209406,-0.02658642,-0.039849266,-0.038377646,0.0024954162,-0.02275275,-0.006569043,-3.5805024e-05,0.005568094,-0.005859618,0.013820019,-0.026902832,-0.008800664,-0.01934489,0.012073304,0.016116058,0.015710963,-0.0037483962,0.004140266,-0.041467804,0.009561584,-0.018550498,-0.004830829,0.00262562,-0.028915724,0.01612334,-0.020695265,0.002773064,0.008701694,0.02337702,-0.0025040272,0.021710323,0.009388487,-0.00068061857,-0.0055366335,0.027267557,0.035958905,-0.0067920676,-0.031524934,0.027056204,0.018050656,-0.025910186,0.023161167,-0.0049318215,0.0052724327,-0.0014541793,0.009687484,0.016734844,0.010902572,0.013239018,-0.017618312,-0.022922985,0.021571683,-7.389846e-05,0.020777646,-0.0105248615,0.006194999,-0.00048764065,-0.008370534,0.012135592,-0.01070121,-0.010396993,0.026869353,0.01757733,0.028068015,0.003107262,-0.023374507,0.0024092663,0.006286342,0.0067184037,0.028109906,0.023908887,0.010466624,-0.027588548,-0.046756785,0.03219264,-0.006893161,-0.015863296,-0.0055272547,0.028391281,0.011975843,0.007451241,-0.015623625,0.0014277311,0.0018540302,-0.021169461,-0.010540996,-0.020193022,0.0039602816,0.01728272,-0.0069098324,-0.010274852,0.003669872,0.001732795,0.01677436,0.010566828,-0.0060203783,0.040264495,0.010831879,0.011330449,0.013708279,0.059379105,0.006981016,0.0055490476,0.002220557,0.055525526,-0.00424103,-0.0031434447,-0.0031981391,-0.018630836,-0.014400827,-0.006273725,-0.0034656643,-0.009612216,-0.010242247,-0.010316715,0.004110621,-0.016647957,-0.015258931,-0.010018563,-0.014475074,0.02867226,-0.0047278437,-0.010294572,-0.015637336,0.010330447,0.014549862,-0.0028305699,-0.017405756,0.037611067,0.03217097,-0.012651611,0.005483631,0.012725593,0.006866217,0.029562796,-0.021368312,-0.012545348,-0.00062706356,-0.000990121,0.03500444,0.012539593,-0.03701506,-7.2406016e-05,0.011917793,0.022847127,0.013725968,-0.041230645,0.024380999,-0.0010587743,0.019998316,-0.032698847,-0.047162715,0.024495576,0.00033526935,-0.00043876836,0.022892362,-0.0035772584,-0.0014317676,-0.0074463557,-0.0053694285,0.010696999,0.025906079,-0.03286149,-0.0025702359,-0.00027340045,-0.027446276,0.0008796983,0.0039796466,0.02427262,0.010161627,-0.01234957,-0.004542698,-0.008168352,0.0049294466,0.0041590505,-0.01635003,0.0031266157,-0.00096267223,0.009691273,0.015175595,0.013698149,-0.0028126598,-0.0038989568,-0.020244755,-0.0050562834,-0.0017858372,-0.029531496,-0.010894113,0.0034024462,0.0013414976,0.008115032,-0.0065365653,-0.015688324,-0.018186698,0.013140493,6.6430155e-05,-0.045489427,0.0031999017,-0.007789527,0.020689948,-0.0114189265,0.0037077197,-0.00956743,-0.003783853,0.024798023,-0.048534743,-0.052032538,0.0053270883,-0.0048445435,0.011272587,0.021356687,0.00442266,0.0017342089,-0.021911006,-0.02693155,0.026659915,0.0025252672,-0.0013049452,0.012139718,-0.010505664,-0.0077691763,-0.01584336,-0.046254933,-0.02603871,-0.021404073,0.021625804,-0.0035035089,0.012403284,0.007094348,-0.016008215,-0.019338187,-0.022899868,-0.005352862,-0.01762802,0.004897159,-0.0062194425,-0.038508467,-0.01970603,-0.014672027,-0.008040195,-0.0014016209,0.016790366,-0.024876608,0.009218823,0.00064470724,-0.022816483,-0.01753112,0.035731327,0.017602194,-0.003121188,-0.027808838,0.0022555464,0.013420234,-0.025657434,-0.019193774,-0.008753231,-0.01502837,-0.018358523,0.002918891,-0.002461211,0.0026942145,0.019822536,-0.020332305,-0.018090185,-0.022003517,0.0066240057,0.015532951,-0.0021284886,-0.0055574407,0.014730119,-0.00044362716,4.517743e-05,0.013646676,-0.01390865,-0.007152868,0.0037624117,-0.0039915154,0.02172917,-0.0020729024,0.0038911784,-0.005897746,0.04875668,0.0009266397,-0.012612268,0.01129631,-0.03158356,0.008507968,-0.005412852,0.0427418,0.007964316,0.00837279,0.00199495,0.020189796,-0.009859358,0.011715596,0.00086466526,-0.014165488,-0.0082370965,0.007391417,0.032357074,-0.023326555,0.0040763747,-0.020743098,-0.00932962,0.0012444854,0.033055358,-0.0058724657,0.010020962,0.017496396,5.9067424e-05,-0.00046409664,-0.030500818,0.0056119263,0.012027289,-0.003950562,-0.017637994,-0.011985323,-0.032737598,-0.025806172,0.0075316094,0.039127093,-0.0064839735,-0.01141306,0.006722649,0.013549414,0.0054634423,0.016067825,0.0058942013,0.041921142,-0.011404866,0.006972154,0.011205888,-0.0011994006,-0.026187664,0.014453912,0.008235175,-0.013944151,-0.0026320496,0.0030676853,0.009007454,-0.023731962,0.030626742,-0.1132691,0.007044357,0.00017170378,0.0044578174,0.024922565,0.008831833,-0.014752311,-0.032135382,0.030434752,0.0005894742,-0.004602342,-0.002843847,0.022988707,0.0012853344,-0.023136823,0.009311112,-0.013499529,-0.04003587,0.037278358,-0.0071469564,0.027787779,0.030506995,0.01587098,0.035924323,0.022259882,-0.026233472,0.0255701,0.00040890122,0.026959147,0.0002913136,-0.028080292,-0.0075953477,0.02007731,-0.013660286,0.030559365,-0.017649004,0.011717115,-0.02256555,0.004084222,-0.022126157,0.016406575,0.01522841,-0.027781216,-0.0012962508,-0.01111162,-0.025208775,-0.009597241,0.014373671,-8.85662e-05,0.01455185,-0.016513092,-0.028588446,-0.010986635,-0.014619462,-0.0036483577,-0.004066838,0.0029673825,-0.0047039893,-0.012905835,0.023484036,-0.035737813,-0.020053085,-0.021632118,0.008453704,-0.01468706,-0.0039479304,0.007883699,0.027692394,0.004450389,0.018931666,-0.015598671,-0.02849282,0.0041672164,0.0035029901,-0.018880187,0.01844196,-0.012023986,0.023567965,0.0065396815,0.009487618,-0.03842466,-0.03315001,-0.053566314,0.00073269394,-0.0139413765,-0.013677172,0.0075769215,-0.0049770977,0.024142463,-0.015035642,-0.014420452,-0.013491386,-0.016467676,-0.019815253,0.039859563,-0.008839409,0.007033792,0.0024466037,-0.00789273,0.0087968055,-0.0037740804,-0.03177923,-0.031218851,0.011049248,-0.012312389,-0.01476991,-0.002180124,-0.011912839,-0.003303158,-0.010066897,0.0034201369,-0.022057623,0.011180911,-0.14716706,-0.0028760452,0.019028358,0.0055344673,0.0146928495,-0.015722377,-0.049377985,-0.012857581,-0.026461711,-0.0050042826,-0.023554653,-0.036916092,-0.016905818,-0.0071881115,0.029551383,0.1310411,0.010869546,-0.010835429,-0.02305054,-0.023160113,-0.015820665,-0.01636609,0.008275749,0.03375024,-0.0032911717,0.00060366764,0.021092253,-0.0077748112,-0.024858985,0.027285803,0.029265838,0.0042194887,-0.0054352623,0.013463688,0.015719937,0.0009895283,-0.015566402,-0.008024426,0.012618142,0.0054336595,0.039803803,0.013061656,-0.0067091202,0.02612918,0.011524716,0.0021363664,-0.016315753,-0.007599532,-0.017233046,0.004104851,-0.013772002,-0.04908563,0.01626915,-0.0066146553,0.014239848,0.009859708,-0.022048509,0.026068207,0.05629273,0.00094075967,0.014169881,0.00076702284,0.017504033,0.016742172,-0.027838167,-0.016399488,-0.009658183,0.009261566,-0.0117806485,0.029872427,0.0071745487,0.02551255,-0.0042003994,-0.0031322795,-0.009243554,-0.006978911,0.0055299522,0.0018614713,0.0088385455,-0.019606546,0.0044398895,0.013422771,0.011694076,-0.034543682,0.008165046,-0.0009767327,0.0055594346,0.011793516,0.014104562,-0.015467103,0.002852202,-0.0034778386,0.013045809,-0.01257457,0.015841154,0.009680688,-0.02076877,0.014374666,-0.004887053,-0.0045377645,0.019987404,-0.032540787,0.0014929153,-0.017530562,0.019538868,0.0013174341,-0.017300172,0.016235603,0.02123478,-0.013747628]",{"tags":37,"relatedLang":48,"relatedPosts":52},[38,40,42,44,46],{"name":16,"slug":39},"megatron",{"name":14,"slug":41},"runtime-scheduling",{"name":15,"slug":43},"readiness-driven-dispatch",{"name":17,"slug":45},"multimodal-training",{"name":13,"slug":47},"pipeline-parallelism",{"id":27,"slug":49,"title":50,"language":51},"rrfp-readiness-driven-pipeline-training-en","RRFP Makes Pipeline Training Follow Readiness","en",[53,59,65,71,77,83],{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":26},"d1c6850c-f832-471b-8beb-c0ebc809667d","peft-bench-fine-tuning-methods-benchmark-zh","PEFT-Bench 讓微調比較更公平","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779179048497-jm5y.png","2026-05-19T08:23:36.803043+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":26},"e24e6e7a-6181-476b-8583-339d854cec68","confident-ai-llm-evaluation-metrics-guide-zh","Confident AI 的 LLM 評估指標指南","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779178456675-x5m6.png","2026-05-19T08:13:46.193772+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":26},"adfa9b15-68b6-44cc-b34d-ebcb02c31210","code-becomes-the-agent-harness-zh","程式碼成了代理引擎","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779173040130-zcyg.png","2026-05-19T06:43:29.625994+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":26},"475844e6-3e2c-49a6-aea0-86a94945d2c2","dashattention-differentiable-adaptive-sparse-attention-zh","DashAttention 讓稀疏長上下文可微","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779171840613-dq1r.png","2026-05-19T06:23:32.886786+00:00",{"id":78,"slug":79,"title":80,"cover_image":81,"image_url":81,"created_at":82,"category":26},"23a3d4c7-5cb7-40ae-a05b-1542364e786f","ibm-prompt-guide-turns-ai-guesses-into-outputs-zh","IBM 提示指南把猜答案變輸出","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779132863293-etob.png","2026-05-18T19:33:55.711767+00:00",{"id":84,"slug":85,"title":86,"cover_image":87,"image_url":87,"created_at":88,"category":26},"7c89c3bd-48cb-4b4e-942d-bbf0409fc392","cattle-trade-llm-bluffing-bargaining-benchmark-zh","Cattle Trade 要測 LLM 談判 bluffing","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779085437419-b0zw.png","2026-05-18T06:23:27.885037+00:00",[90,95,100,105,110,115,120,125,130,135],{"id":91,"slug":92,"title":93,"created_at":94},"f18dbadb-8c59-4723-84a4-6ad22746c77a","deepmind-bets-on-continuous-learning-ai-2026-zh","DeepMind 押注 2026 連續學習 AI","2026-03-26T08:16:02.367355+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"f4a106cb-02a6-4508-8f39-9720a0a93cee","ml-papers-of-the-week-github-research-desk-zh","每週 ML 論文清單，為何紅到 GitHub","2026-03-27T01:11:39.284175+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"c4f807ca-4e5f-47f1-a48c-961cf3fc44dc","ai-ml-conferences-to-watch-in-2026-zh","2026 AI 研討會投稿時程整理","2026-03-27T01:51:53.874432+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"9f50561b-aebd-46ba-94a8-363198aa7091","openclaw-agents-manipulated-self-sabotage-zh","OpenClaw Agent 會自己搞砸自己","2026-03-28T03:03:18.786425+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"11f22e92-7066-4978-a544-31f5f2156ec6","vega-learning-to-drive-with-natural-language-instructions-zh","Vega：使用自然語言指示進行自駕車控制","2026-03-28T14:54:04.847912+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"a4c7cfec-8d0e-4fec-93cf-1b9699a530b8","drive-my-way-en-zh","Drive My Way：個性化自駕車風格的實現","2026-03-28T14:54:26.207495+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"dec02f89-fd39-41ba-8e4d-11ede93a536d","training-knowledge-bases-with-writeback-rag-zh","用 WriteBack-RAG 強化知識庫提升檢索效能","2026-03-28T14:54:45.775606+00:00",{"id":126,"slug":127,"title":128,"created_at":129},"3886be5c-a137-40cc-b9e2-0bf18430c002","packforcing-efficient-long-video-generation-method-zh","PackForcing：短影片訓練也能生成長影片","2026-03-28T14:55:02.688141+00:00",{"id":131,"slug":132,"title":133,"created_at":134},"72b90667-d930-4cc9-8ced-aaa0f8968d44","pixelsmile-toward-fine-grained-facial-expression-editing-zh","PixelSmile：提升精細臉部表情編輯的新方法","2026-03-28T14:55:20.678181+00:00",{"id":136,"slug":137,"title":138,"created_at":139},"cf046742-efb2-4753-aef9-caed5da5e32e","adaptive-block-scaled-data-types-zh","IF4：神經網路量化的聰明選擇","2026-03-31T06:00:36.990273+00:00"]