Xiga xigaHagaha xiga
Isku-dheelitirka LLM iyo Isku-dheellitirka Culayska
Farsamo
HAGAHA Farsamada
Gating iyo marin-u-dhigiddu waxay u oggolaadaan shabakada neerfaha inay hawlgeliso oo keliya qaybaha ay u baahan tahay talo-gelin kasta halkii ay ka shaqayn lahayd moodelka oo dhan mar kasta.
Dhageyso hagahanAI-waa laga sheekeeyay · ~2 min
Tani waxay kala saartaa cabbirka moodeelka laga soo bilaabo kharashka xisaabinta, taas oo awood u siinaysa moodooyinka aadka u weyn ee si degdeg ah u socda oo raqiis ah si ay u ordaan.
Xisaabinta shuruuda ah macnaheedu waa shabakadu waxay samaysaa go'aamo ku-tiirsanaan xog ku saabsan hab-hoosaadyada la isticmaalayo. Shabakad yar oo 'gating' ama 'router' ah oo la bartay ayaa eegtaa gelinta kasta (badanaa calaamad kasta) waxayna soo saartaa buundooyinka xulashada 'khubarada' loo diro. Lakabka Isku-dhafka-Khubarada (MoE), daraasiin ama boqollaal shabakad-hoosaadyo khabiiro ah ayaa jira, laakiin router-ku wuxuu soo xushaa mid ama laba sare oo keliya calaamo kasta, sidaa darteed khubarada intooda badani waxay joogaan shaqo la'aan wax kasta oo la soo galiyo. Natiijadu waa moodel leh wadarta guud ee tirinta cabbirka laakiin tiro yar oo firfircoon, taasoo siinaysa awoodda matalaadda moodeelka weyn ee qiimaha runtime ee mid aad uga yar. Tani waa sida moodooyinka ay ka midka yihiin Transformer-ka Beddelka, GLaM, iyo noocyo badan oo luuqado waaweyn oo xuduud ah oo cabbiraya trillions oo cabbirro ah oo la awoodi karo.
Beddelka Transformer-ka waxa uu ku hagayaa calaamad kasta hal khabiir, isaga oo u miisaamaya in ka badan trillion xuduudo iyada oo la ilaalinayo midkiiba mid hoose.
Moodooyinka luqadaha waaweyn ee xudunta ah ee isticmaalaya lakabyada Isku-dhafka-Khubarada sidaa darteed qayb yar oo miisaanno ah ayaa dhaqaajiya calaamad kasta.
Kala soocida sawirka hore-u-soo-joojinta lakabka gacmeed ee sawirada fudud oo si qoto dheer u socda kaliya kuwa adag.
Moodooyinka luqadaha badan ee router-koodu baranayaan inay ka soo diraan calaamado luqado kala duwan khubaro gaar ah oo kala duwan.
Gating yar-yar ayaa hadda udub dhexaad u ah cabbirida moodooyinka xuduudaha, iyo isbeddelku waxa uu ku wajahan yahay khubarada fir-fircoon, dariiqyada xariifnimada leh, iyo ku-wareejinta lakabyo badan. Filo farsamooyin ka wanaagsan tababarka xasiloon, hoos u dhaca isgaadhsiinta marka khubaradu ay ku faafaan dardar-geliyayaal badan, iyo falanqaynta 'khabiir takhasuska' si loo fahmo waxa khabiir kastaa barto. Xisaabinta shuruuda ah waxay sidoo kale ku faaftaa wixii ka baxsan MoE una gudubta shabakadaha ka bixida hore iyo moodooyinka qoto dheer ee firfircoon kuwaas oo ku kharash gareeya kaliya agabyada adag.
Warqad bilaash ah
Saddex sheeko oo la xaqiijiyay oo AI ah maalin kasta oo shaqo ah, oo lagu qoray Ingiriis fudud. Bilaash had iyo jeer, aan xayaysiis lahayn.
Hal iimayl maalin kasta oo shaqo ah. Ka bax hal gujin. Marna ma iibinno ama ma wadaagno ciwaankaaga.
Is tijaabi
Fikrad degdeg ah oo ku saabsan jawaab kasta, iyo shahaado la wadaajin karo oo leh aqoonsi la xaqiijin karo markaad ka gudubto koorso.
Taageer waxbarashada AI bilaashka ah. AI Understanding waa 501(c)(3) aan faa'iido doon ahayn — xayaysiis la'aan, albaab lacag ah ma jiro, had iyo jeer. Samee deeq
Gating iyo marin-u-dhigiddu waxay u oggolaadaan shabakada neerfaha inay hawlgeliso oo keliya qaybaha ay u baahan tahay talo-gelin kasta halkii ay ka shaqayn lahayd moodelka oo dhan mar kasta. Tani waxay kala saartaa cabbirka moodeelka laga soo bilaabo kharashka xisaabinta, taas oo awood u siinaysa moodooyinka aadka u weyn ee si degdeg ah u socda oo raqiis ah si ay u ordaan.
The guide defines data-dependent selection of submodules as conditional computation.
The guide says a learned router selects experts based on its input.
Sparse routing can activate only part of the total parameter set per input.
The guide explains that most experts can remain idle for a particular token.
The guide notes routers can favor a few experts, leaving others underused.
Sii wad waxbarashada
Tilmaamayaal badan ayaa loo doortay mawduucan
Xiga xigaHagaha xiga
Isku-dheelitirka LLM iyo Isku-dheellitirka Culayska
Farsamo