RWKV Lineer
RWKV architecture la buy tàggat ni Transformer waaye di doxal inference ni reso buy baaxoo, di joxe temps ligneer, defar mémoire buy wéy.
Résumé
It reformulates attention so there is no quadratic cost and no growing key-value cache.
Plongeur bu xóot
RWKV (ñu koy waxe 'RwaKuv') mooy Tekki, Diisaay, Caabi, Valeer, ñeenti mbir yu am solo yi. Ñu ngi ko sos muy projet bu ubbeeku, bu askan wi taxawal, Bo Peng moo ko jiite. Luñu bëgga mooy ñu mëna tàggat Transformers ci anam wu méngoo, boole ci dindi seen inference bu seer. Standard attention dafay denc cache bu am valeur key buy màgg ak token bu nekk ba noppi di méngale token bu bees bu nekk ak yeneen yi ko jiitu. RWKV lu moy loolu dafay yóbbaale benn stade bu nëbbu bu ndaw, diko yeesal ak sàrtu time-decay suko defee xibaar bu yàgg bi di dem bu baax. Ci diiru tàggat yaram mën nañu ko dindi ci anam wu ñuy paralel; ci jamonoy generation dafay nekk RNN di defar benn token benn yoon ci njëg bu dul soppeeku. Loolu moo tax mu neex ci contexte yu gudd ak jëfandikoo jumtukaay yu néew.
Gis-gis xarala
RWKV dafay wecci softmax dot-produit bi ak xeetu ligneer buy baamtu. Benn poids de decay ci chaine bu nekk (W) mooy saytu ni caabi yi weesu di gaawee ñàkka am doole, buntu biy jot (R) mooy dogal ba ñaata stade accumulé lañu wara jàng, ba noppi vecteur key/valeur yi dañuy dundal somme pondérée biy dawal. Ndax jéego bu nekk mingi aju ci nekkin bi njëkk, mémoire bi dafay wéy di nekk te liggéey bi token bu nekk du màgg ak guddaayi sequence bi.
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 RWKV Linear
RWKV dafa baamtu ci anam wu gaaw ci ay xeetu (v4, v5 Eagle, v6 Finch ak ginaaw), wàññi gap bi am ci kalite ak Transformers boole ci tëye njëg ligneer. Xaarandil yokkute bu wéy ci model yu ubbeeku ci làkk yu bari, edge ak CPU deployment fu memory bu sax bi am solo, ak jëmmal hybrid. Inference bimu àndal yépp moo tax mu nekk kàndidaa bu am doole ci aplikaasioŋu streaming ak ci contexte yu gudd lool fu cache yu am valeur bu am solo di kalaate.
Doxal ci àdduna dëgg
Doxal xeetu waxtaan yu mëna ubbeeku ci CPU yi wala aparey yu am mémoire bu néew ak mémoire budul soppeeku ci token bu nekk
Defar mbindu streaming fu ñuy defaree ay token benn-benn te du am benn cache buy màgg
Liggéeyukaay dokimaa yu gudd fu cache valeur-caabi bu Transformer di nekk lu yaatu lool
Dëkkandoo yi ak projet yu bari lakk yu soxla architecture bu baax te am lisaas bu ubbeeku
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 RWKV Linear Attention 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
Lineer ak Kernel yiy def
Laaj yi ñuy faral di laaj
What is RWKV Linear Attention?
RWKV architecture la buy tàggat ni Transformer waaye di doxal inference ni reso buy baaxoo, di joxe temps ligneer, defar mémoire buy wéy. Dafay formuleewaat bàyyi xel suko defee amul benn njëg bu kwadratik te amul benn cache bu am valeur key buy màgg.
Lan mooy moomeel bu mag bi ci RWKV?
RWKV's mébetu jëmmal mooy tàggat yaram bu nuroo ak Transformer boole ci inference buy baaxoo am.
Lu arafu RWKV yi di tekki?
RWKV ñu ngi ko tuddee ci ñeenti mbir yi ci bokk: Reception, Poids, Key, ak Valeur.
naka lay def ba RWKV mëna wéy di am mémoire ci diiru generation bi?
RWKV melni RNN, dafay yeesal jéego bu nekk ci dayo bu takku, suko defee mémoire bi du màgg ak guddaayu toppalante bi.
Ban cër la poids de décay temps (W) di def?
Diisaayu decay biñu jàngee ci chaine bu nekk mooy wane ni butoŋu past yi di gaawe fade ci stade biy daw.
Buñu ko méngale ak bàyyi xel ci softmax, lan la RWKV di moytu ci bàyyi xel ci ligneer?
Soo jëfandikoo recurrence, RWKV dafay moytu méngale bépp token ak bépp token, dindi njëgu quadratic bi.