Kwenzekeni
Kokuthunyelwe kobunjiniyela kwangomhla zingama-28 ku-Agasti 2026, i-Databricks yachaza i-AI Runtime API ngokwenza imisebenzi emikhulu yokuqeqesha ye-PyTorch ikwazi ukumelana nokwehluleka kwe-GPU kanye namapayipi okufakwayo ahamba kancane. Inkampani incoma ukubhekwa okusatshalaliswayo, ukonga okuvumelanayo, ukutholwa okuzenzakalelayo, ukugcinwa kwesikhashana kwendawo kanye nokulanda kuqala, kanye nokuhlola ipayipi ledatha kanye nesimo sejeneretha yenombolo engahleliwe eceleni kwezisindo zemodeli.
Okuthunyelwe bese kuncoma ukusebenza kokulondoloza okungavumelani kwe-PyTorch. Kulo mklamo, ukuqeqeshwa kukhokhela ikhophi esheshayo ibe isilondolozi sesiteji kuyilapho ukulayisha kuqhubeka ngemuva. I-Databricks ithi UCVolumeWriter ye-AI Runtime kanye ne-UCVolumeReader zisebenzisa isiteji se-NVMe yendawo futhi zimake indawo yokuhlola njengeqediwe kuphela ngemva kokuba yonke idatha isifinyelele lapho iya khona. Ngakho-ke le ndlela ihlukanisa iphuzu lapho ukuqeqeshwa kungaqhubeka khona kusukela ekuphothulweni kwakamuva komsebenzi wokulondoloza, kuyilapho sisabophela ukuqedwa kwendawo yokuhlola endaweni kunokuba nje kuyikhophi yendawo. Lo mehluko ubalulekile encazelweni yokuthunyelwe yokuqeqeshwa okubekezelela amaphutha kanye nokugeleza komsebenzi wokubuyisela.
Ekuqhathanisweni okubikwe yinkampani, imodeli yolimi ye-DDP engu-2.8-billion- kuma-32 H100 GPUs ithathe amasekhondi angu-36 nge-async_save uma iqhathaniswa namasekhondi angu-66 nge-torch.save, noma izikhathi ezingu-1.8 ngokushesha. Ngemodeli yepharamitha yebhiliyoni engu-20 kuma-GPU angu-32 H100, ithebula libika amasekhondi angu-9 uma kuqhathaniswa namasekhondi angu-522, noma izikhathi ezingu-58 ngokushesha. Lezi zibalo zichaza ukuqhathanisa kwesikhathi sokuhlola okwethulwa yi-Databricks kumamodeli ashiwo nehardware. Anikezwa njengobufakazi bevelu yokulondoloza okuvumelanayo, kuyilapho ehlala eboshelwe ekucushweni nasemongweni wokulinganisa ochazwe kokuthunyelwe.
Okuthunyelwe kuthi ukuqhathanisa akubandakanyi isikhathi sokulondoloza inethiwekhi se-torch.save. Lezo ziqu zichaza lokho okwenziwa isiqhathaniso sesikhathi esibikiwe futhi esingakuhlanganisi. Isincomo ngenxa yalokho sibanzi kunesibalo esisodwa sesivinini: indawo yokuhlola esabalalisiwe, ukonga ingemuva, ukutholwa okuzenzakalelayo, isiteji sendawo, ukulondoloza isikhashana kanye nokulanda ngaphambili kwethulwa njengezingxenye ezihlobene zomklamo wokuqina. Ngokuhlangene, imininingwane ibonisa ukuthi i-Databricks ixhuma kanjani imishini yokuhlola nenkinga ebonakalayo yokugcina umsebenzi omkhulu wokuqeqesha uhamba ngemva kokuphazamiseka noma ukulethwa kokufakwayo okunensayo, ngaphandle kokushintsha izilinganiso ezibikwe emthonjeni.
Imininingwane yomthombo: databricks.com ↗
Kungani kubalulekile
Ukugijima okukhulu kokuqeqeshwa kwe-AI kungamosha isikhathi esikhulu sokusheshisa lapho umsebenzi wehluleka, ulinda idatha, noma uqala kabusha endaweni engafanele kudathasethi. I-Databricks yethula izilinganiso ezithile ezisikisela ukuthi indlela yayo ingathuthukisa isikhathi sokuhlola kanye nokuphuma kokuqeqeshwa kwesithombe, nakuba izibalo zibikwe yinkampani futhi zincike kuhadiwe elihloliwe, umthwalo womsebenzi nendlela yokugcina.
I-Databricks iphinda ihlonze ingozi yokunemba okungenzeka ingakhiqizi ukwehluleka okusobala. Uma umsebenzi ulondoloza imodeli, isilungiseleli kanye nesinyathelo sokuqeqesha kodwa hhayi indawo yesilayishi idatha, ukuqalisa kabusha kungaphinda izibonelo esezibonile futhi kweqe izibonelo ebezingakacutshungulwa. Okuthunyelwe kuthi ukuqalisa kabusha okuphindaphindiwe kungase kuguqule ukusatshalaliswa kwedatha okusebenzayo ngaphandle kokukhiqiza iphutha. Ngakho-ke ukukhathazeka kumayelana nezinqubo zomsebenzi oqaliswe kabusha, hhayi kuphela ukuthi umsebenzi uqala kabusha ngempumelelo. Indawo yokuhlola ingabonakala isebenziseka kuyilapho ubudlelwano phakathi kwesimo sokuqeqeshwa esilondoloziwe nendawo yedatha bungaphelele.
Amakhambi ayo ahlongozwayo afaka isampula yokurekhoda noma ama-shard offset, ukuma kwedathasethi ye-serializing, noma ukukhomba emingceleni yenkathi. Lezi zixazululo zibhekana nolwazi lwendawo olungekho oluchazwe kokuthunyelwe ngokwenza ipayipi ledatha libe yingxenye yesimo esibuyiselekayo. Ukukhetha phakathi kwazo kuhlala kungaphakathi kwesiqondiso sokuqaliswa esethulwa yi-Databricks, futhi imfuneko eyisisekelo ukuthi umsebenzi oqaliswe kabusha ugcina ubudlelwano obuhlosiwe phakathi kwenqubekelaphambili yokuqeqeshwa nenqubekelaphambili yedathasethi. Kungakho okuthunyelwe kuphatha ukukhomba ipayipi ledatha njengengxenye yokulunga kokuthola kabusha esikhundleni semininingwane yokusebenza kokuzithandela.
Iphinde ithi imbewu yokushova nokwengeza kanye nezindawo ezikhiqiza izinombolo ezingahleliwe kumele zilondolozwe ukuze i-oda ledatha eliqaliswe kabusha lihlale likwazi ukukhiqizwa kabusha. Ukulondoloza lezo zifunda kunweba indawo yokuhlola ngale kwezisindo zemodeli, ulwazi lwe-optimizer nesinyathelo sokuqeqesha. Kufreyimu yomthombo, ukukhiqiza kabusha kuncike ekulondolozeni indawo yesilayishi sedatha kanye nesimo esilawula uku-oda nokwengeza. Okushiwo okungokoqobo ukuthi ukutakula kufanele kuhlolwe kokubili ukuqhubeka nokunemba: umsebenzi kufanele ubuyele esimweni esisebenzisekayo, futhi ukucutshungulwa okuqaliswe kabusha kufanele kubonise isimo obekuhloswe ngaso ukusindiswa.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ukulandelela okubalulekile ukuthi ingabe le miphumela ibambe ngaphandle kokucushwa kwe-Databricks nokuthi ingabe ama-API ayatholakala yini ngokuhambisana okucacile, amanani kanye neziqondiso zokusebenza. Abasebenzisi kufanele futhi babheke izivivinyo ezizimele zokulunga kokutakula, ukuqina kwendawo yokuhlola, ukukhulisa usayizi weqoqo kanye nokugcinwa kwe-oda ledatha, hhayi nje ukugijima kwebhentshimakhi okusheshayo.
I-Databricks ithi amarekhodi ayo e-DataLoader fetch_seconds ku-MLflow, enikeza opharetha indlela yokuhlonza amaqoqo ashiya ama-GPU elindile. Leyo methrikhi ingenza isistimu ixilongwe kalula, kodwa yona ngokwayo ayisunguli inani eliphansi lezindleko noma ikhwalithi engcono yemodeli. Ihlinzeka ngokuqaphela mayelana nesikhathi sokulanda kanye nokulinda kwe-GPU okungenzeka kube khona, kuyilapho umphumela omkhulu uncike kuwo wonke umzila wokuqeqesha nowokugcina. Ngakho-ke imethrikhi iwusizo njengesignali yokusebenza ngaphakathi kwesistimu ehlongozwayo, kodwa ayethulwa njengesilinganiso esiphelele senani lesistimu.
Abasebenzisi bazodinga ukubalwa kwezimali kokuphela kuya ekupheleni okuhlanganisa umthamo wendawo we-NVMe, ukufudumala kwenqolobane, ukudluliswa kwenethiwekhi, izindleko zokulondoloza, imvamisa yomsebenzi ohlulekile kanye nezindleko zokubala zanoma iyiphi idatha yokuqeqeshwa eyimpinda noma eshintshiwe. Lokho kucatshangelwa kuxhumanisa ingxoxo yokusebenza nenkinga yokunemba echazwe ngenhla. Umsebenzi wokuhlola osheshayo noma ukubambezeleka kokufakwayo okubonakalayo ngeke, ngokwako, kuxazulule imibuzo emayelana nezinsiza, ukuziphatha kokutholwa kanye nokuphathwa kwedatha. Ukubalwa kwezimali okuceliwe kufanele ngenxa yalokho kukhave ukugeleza komsebenzi okugcwele okumelelwe emthonjeni, okuhlanganisa izindleko nemithelela ehlala ngaphandle kwezikhathi zebhentshimakhi ngayinye.
Umthombo unikeza ithisisi eqinile yokusebenza kanye nesiqondiso sokusetshenziswa esiwusizo, kuyilapho ushiya lokho kuqhathanisa okubanzi kungaziwa. Ukulandelela okubalulekile ngakho-ke ukuhlola ukucushwa okushiwo, ukutholakala nezimo zokusebenza ezihambisana nezilinganiso ezibikiwe. Abasebenzisi kufanele babheke izivivinyo ezizimele zokulunga kokutakula, ukuqina kwendawo yokuhlola, ukukhulisa usayizi weqoqo kanye nokugcinwa kwe-oda ledatha, hhayi nje ukugijima kwebhentshimakhi okusheshayo. Lokho kuhlola kungasiza ukuthola ukuthi ingabe ama-API achaziwe aletha ukuziphatha okufanayo ngale kwemibiko ye-Databricks yokulungiselelwa, kuyilapho kugcinwa ukungaqiniseki okukhonjwe emthonjeni.