[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82674-en":3,"doc-seo-82674-105":29,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},82674,3848291630094,"Emma Wilson","https://eur-avatar.wpscdn.com/davatar_085a072bc5b1113ac321206ff7593b45",6,"Technology","Not Every Sync Is Safe: Calibrated DiLoCo Scheduling for Shared AI Infrastructure","DiLoCo-style training reduces communication by letting learner islands optimize locally before occasional outer synchronization, which suits fragmented industrial AI fleets that share hardware with latency-sensitive serving. The core issue is when an outer merge is worth its cost, and whether deferring particular windows changes outcomes. Existing work often compares workload-aware policies to fixed cadences while missing budget-isolated control. Matched-random deferral neutralizes placement bias, revealing that only calibration plus burst-aware forecasting justifies serving-SLO improvements.","Not Every Sync Is Safe: Calibrated DiLoCo Scheduling for Shared AI  \nInfrastructure  \nMaxwell Twelftree1 , David Lemphers1 , An-chi He1 , Yue Yang1 ,  \n1Maincode, Melbourne, Australia,  \nCorrespondence: yue@maincode  \narXiv :2607 .02544v1 [ cs .DC] 24 Jun 2026  \nAbstract  \nDiLoCo-style training reduces communication by letting learner islands train locally before occasional outer synchronization, making it attractive for fragmented industrial AI fleets where training shares hardware with latencysensitive serving. The question for such fleets is when an outer merge is worth its system cost, and whether choosing which windows to defer matters at all. Existing scheduling studies evaluate workload-aware policies against fixed-period baselines, but most omit the control that isolates timing from budget: matched random deferral, which inherits the controller’s synchronization budget but is not itself deployable. This omission is consequential: across controlled stress tests and real vLLM sidecar replays, matched random ties or beats every forecast-free policy we test, so gains reported against weaker baselines cannot be attributed to window choice. We fill this gap with Workload-Aware DiLoCo (WA-DiLoCo), a score-based controller that weighs learner progress against fleet pressure, and a calibration protocol that determines when matched random can be beaten, then demonstrate that it can. In the bursty regime where calibration exposes request-overlap structure, adding a onestep EWMA burst forecast to the online controller beats matched random in real vLLM  \nsidecar replay, reducing SLO violations from 6.54% to 5 .09%(8 of 10 seeds, p = 0 .021); offline Calibrated-WA, a non-deployable bound, shows the remaining headroom at 4.45% versus 6.26% . The deployable lesson remains the protocol: report real-sidecar effect-size transfer, ano-sync load match, and a matched-random envelope before claiming serving-SLO improvement.  \n1 Introduction  \nConventional distributed training assumes a tightly coupled cluster in which thousands of GPUs synchronize gradients at high frequency over expensive interconnects. Real industrial environments  \nrequest bursts outer merge (∼8 s)  \nFixed-H (DiLoCo default)  \n3/3 in bursts  \nMatched random (control, same budget) 2/3 this draw  \nWorkload-aware + forecast (ours) 0/3  \ndeferred past burst  \ntime  \n merge in burst → SLO violations  merge in valley → cheap  \nFigure 1: The scheduling problem. An outer merge occupies the fleet for ∼8 s, and any request overlapping that window inherits the interference. Each row places the same budget of three merges: a fixed cadence collides with bursts, matched random (one draw) is equally blind, and a workload-aware policy defers merges into valleys. Placement is the only lever matched random does not capture; calibration (§4) decides when it is worth pulling.  \nrarely satisfy this assumption: compute is fragmented across clusters, regions, cloud providers, hardware generations, and availability windows.  \nDiLoCo-style training reduces communication by allowing learner islands to perform many local optimization steps before an outer synchronization (Douillard et al., 2023), which makes it attractive for these fragmented fleets. However, reducing how often synchronization occurs does not answer an equally important production question: when is it safe to synchronize? Recent DiLoCo variants make synchronization cheaper, less blocking, and more failure-tolerant (Douillard et al., 2025, 2026 ; Koneputugodage et al., 2026 ; Kolehmainenet al., 2025 ; Sarfi et al., 2025), but they do not decide whether the current serving state makes an outer merge safe. This question becomes critical in shared fleets, where an outer merge can consume host, network, and checkpoint resources at exactly  \nthe moment latency-sensitive serving is under pressure; a fixed interval ignores fleet state, a pressure gate may defer progress too aggressively, and no single policy should be expected to dominate ","cbCaigmNB92jTXSo","https://ap.wps.com/l/cbCaigmNB92jTXSo","pdf",400313,1,16,"English","en",105,"# Abstract\n# Introduction","[{\"question\":\"What problem does the paper address in shared AI infrastructure?\",\"answer\":\"It addresses deciding when outer synchronization (outer merge) is safe in fleets where training resources contend with latency-sensitive serving, and where request bursts can cause interference.\"},{\"question\":\"What is matched random deferral and why is it important?\",\"answer\":\"Matched random deferral is a control that uses the same synchronization budget as the tested scheduler but randomizes placement. It isolates the effect of choosing windows, ensuring reported gains are not incorrectly attributed to placement.\"},{\"question\":\"How does the proposed method improve SLO outcomes in bursty regimes?\",\"answer\":\"Workload-Aware DiLoCo with a one-step EWMA burst forecast beats matched random in real vLLM sidecar replay, reducing SLO violations from 6.54% to 5.09% (8 of 10 seeds, p=0.021).\"}]",1784182213,40,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"not-every-sync-is-safe-calibrated-diloco-scheduling-for-shared-ai-infrastructure","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/technology/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/not-every-sync-is-safe-calibrated-diloco-scheduling-for-shared-ai-infrastructure/82674/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does the paper address in shared AI infrastructure?","Question",{"text":75,"@type":76},"It addresses deciding when outer synchronization (outer merge) is safe in fleets where training resources contend with latency-sensitive serving, and where request bursts can cause interference.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What is matched random deferral and why is it important?",{"text":80,"@type":76},"Matched random deferral is a control that uses the same synchronization budget as the tested scheduler but randomizes placement. It isolates the effect of choosing windows, ensuring reported gains are not incorrectly attributed to placement.",{"name":82,"@type":73,"acceptedAnswer":83},"How does the proposed method improve SLO outcomes in bursty regimes?",{"text":84,"@type":76},"Workload-Aware DiLoCo with a one-step EWMA burst forecast beats matched random in real vLLM sidecar replay, reducing SLO violations from 6.54% to 5.09% (8 of 10 seeds, p=0.021).","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,113,117,122,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":111,"slug":112},50,"technology",{"id":114,"doc_module":4,"doc_module_name":45,"category_name":115,"show_sort_weight":28,"slug":116},7,"Healthcare","healthcare",{"id":118,"doc_module":4,"doc_module_name":45,"category_name":119,"show_sort_weight":120,"slug":121},8,"Research & Report",30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]