Njaxasu xóotaayu
Njaxasu xóotaayu (MoD) dafay may transformatër bi mu jëfandikoo ay lim yu wuute ci ay jeton yu wuute, di yóbbu jeton yu 'am solo' yi ci xayma yu diis yi ci layer bu nekk.
Résumé
It cuts the cost of processing easy tokens while keeping a fixed, predictable compute budget.
Plongeur bu xóot
Transformatër yiñ miin dañuy jëfandikoo bepp couche ci bepp token, ba ci yu tuuti yu melni ponctuation. Njaxasu xóotaayu, bi Google DeepMind dugal ci 2024, dafay yokk benn routeur bu ndaw ci blok bu nekk buy tànn benn fraction top-k bu fixe ci token yi ngir mëna def seen xel ci seen bopp ak xayma MLP; ñi ci des dañuy romb blok bi jaaraleko ci lëkkalekaay bu des. Ndax k jetons kese lañuy liggéey ci layer bu nekk, total compute (FLOPs) dañu koy tënk te xam ko lu jiitu, wuute na ak pexe yu njëkk ya dynamic-depth yu wuute ci anam wu ñu mënul xam. Loolu mooy tax batching bi ak jëfandikoo hardware bi gëna am njariñ. Mod-trained models mën nañu méngoo ak kalite transformateur baseline jëfandikoo FLOPs yu néew ci pass forward bu nekk, wala yegg ci kalite bu gëna kawe ci benn ordinatër, te xalaat bi dafay boole ci boppam ak Mixture-of-Experts ngir joxe 'MoDE' models yiy jaar ci xóotaayu ak yaatuwaay.
Gis-gis xarala
Ci bloc MoD bu nekk, routeur lineaire bu jàng dafay poñ token bu nekk ba noppi tëye top-k ci poñ; jetons yiñ tann ñu ngi jaar ci attention ak MLP, waaye jetons yiñ tannul ñu ngi leen di yóbbu te duñu soppiku ci yoon wi des. Jëfandikoo top-k bu takku (te du benn threshold ci token bu nekk) dafay tax graphique bi nekk static ak forme tensor yu fippu, te loolu lu baax la ci hardware bi. Routeur bi dañu ko tàggat ak yeneen reso yi, ba noppi generation causal dafay jëfandikoo ay predictor yuy jàppale suko defee dogal yi ñuy jël ñu baña xool ci token yi ci kanam.
njeextalu pexe
Gaawaay ak yaatuwaay
Liggéeyukaay yi ci làkk yi mën nañu gëna gaaw te duñu yàq deggoo gi.
Dugg ak yegg
Dafay yaatal jëfandikoo gi ci làkk yi ak ci anam yi ñuy jokkoo.
dogal yu gëna leer
Ekip yi mën nañu gëna yàgg ci àtte ci jamono ji otomatisation di liggéey ci baamtu.
Ëlëgu njaxasu xóotaayu
calcul conditionnel levier bu mag la ci efficacité ni model scale, te MoD misaal la bu teel, sell. Xaarandi lëkkaloo bu gëna xóot ak Mixture-of-Experts (routing ci xóotaayu ak eksper yi), budget yiy méngoo ak li ñuy gëna xëcc ngir dugal ay done yu yomb, ak routeurs yu jàng yu gëna mëna xàmmee ban token moo soxla liggéey bu xóot. Ginaaw njëgu inference mooy ëpp doole ci koom-koomu jëfandikoo gi, pexe yiy may model yi ñu 'gëna xalaat' ci barab yiñ soxla, boole ci tëye latency buñ mëna seentu, mën nañu nekk standard ci architecture yu mag yi.
Doxal ci àdduna dëgg
Wàññi FLOPs yiñ soxla ngir liggéey ci këyitu gudd yi ci sànni xayma yu xóot yi ci token yu fees
Taggat xeetu liggéey bu méngoo ak kalite bu njëkk bi ci ordinatër bu woyof, wàññi njëgu liggéey bi
boole ak njaxasu-ekspert (MoDE) ngir jaar ci xóotaayu couche ak tànneef eksper
Fexe bañu mëna seentu, latency bu takku ci jeton bu nekk ndax budget bi ci layer bu nekk dañu ko tëral bu njëkk
Risk yi ak balustrade yi
Lépp lu jaarul yoon mën na dugg ci rapoor yi, jàppale ci liggéey bi, wala ci njariñu gëstu bi.
Sensibilite bu gaaw mën na jur njariñ yu wuute ci laajte yu noonu mel.
Done yu am solo mën nañu feeñ sudee seytu jëfandikoo gi néew doole.
Roadmap ngir samp gi
Mandargal formaa génne gi, melokaan bi, ak standard kalite yi laata ngay dugal ko.
Tontu yu am solo ak balluwaay yu wóor saa yu dëggu bi di am solo.
Fexeel am barabu xool nit ñi ngir am njariñ yu am solo.
Toppal anami gacce yi ak di faral di tàggataat ay laaj wala def-liggéey.
Weyal di banneexu
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Mixture of Depths quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Gis bi ci topp
Aggrégation njaxasu Agent
Laaj yi ñuy faral di laaj
What is Mixture of Depths?
Njaxasu xóotaayu (MoD) dafay may transformatër bi mu jëfandikoo ay lim yu wuute ci ay jeton yu wuute, di yóbbu jeton yu 'am solo' yi ci xayma yu diis yi ci layer bu nekk. Dafay wàññi njëgu liggéey token yu yomb yi boole ci tëye budget ordinatër buñ mëna seentu.
Lan mooy wuute ci njaxasu xóotaayu jeton yi?
MoD yenn jetons rek lay jaar ci calcul bu diis bu layer bu nekk, kon jetons yu wuute yi dañuy am xóotaayu wuute.
naka la ko MoD di def ngir mëna xam bu baax budget bi?
Tann top-k bu fixe ci jetons ci block bu nekk dafay tëj FLOPs te dafay tëye forme tensor yi, te loolu lu baax la ci hardware bi.
Lan mooy dall token yi routeur bi tannul ci benn block buñ jox?
Jeton yiñ tannul dañuy romb xayma blok bi ba noppi ñu yóbbu leen kanam te duñu soppiku ci yoon wi des.
Kan moo dugal njaxasu xóotaayu?
DeepMind moo dugal njaxasu xóotaayu mbir yi ci benn téere ci atum 2024 buy wax ci xayma transformatër buy soppiku.
Lan mooy jur boole MoD ak njaxasu-ekspert yi?
Njaxas xóotaayu yoon wi ak yoon wu xarañ yi dafay jur xeetu 'MoDE' yuy indi xayma ci diisaay yi ak xarañ yi yépp.