Kunyerera Idirishya
Kunyerera idirishya ryitondewe rigabanya buri kimenyetso cyo kwitabira gusa mubunini-bunini buturanye bwibimenyetso byegeranye aho kugirango bikurikirane.
Incamake
This cuts the quadratic cost of standard attention down to linear, making long-context models far cheaper to run.
Kwibira cyane
Kwiyitaho bisanzwe kugereranya buri kimenyetso nikindi kimenyetso, bityo rero uburebure bwa N busaba kugereranya N-kwaduka. Kunyerera idirishya ryitondewe bikosora ibi muguha buri kimenyetso idirishya ryubunini W (vuga ibimenyetso 4.096) kandi witabire abaturanyi gusa muriryo dirishya. Igiciro gikura nka N inshuro W aho kuba N-kare. Byibanze, gutondekanya ibice byinshi byamadirishya byagura umurima wakira neza: nyuma ya L, amakuru arashobora gukwirakwira hafi ya L inshuro W ibimenyetso, nkumurima wa CNN ukura. Mistral 7B yamenyekanye cyane hamwe nidirishya rya 4.096-ryerekana hejuru ya 32, igera kuri theoretical 131K-token span. Abanyamideli bakunze kuvanga idirishya ryibice hamwe na rimwe-byuzuye byuzuye kugirango babungabunge intera ndende.
Ubushishozi
Muri mask yo kwitondera, ikibazo kumwanya i nemerewe gusa kubona urufunguzo ruva kumwanya i gukuramo W wongeyeho 1 kugeza i (urubanza rwimpamvu). Iyi maskike isobanura KV cache ikenera gusa ibimenyetso bya W byanyuma kuri layer, kugabanya kwibuka mugihe cyibisekuru. Kuberako idirishya rihinduranya na buri kimenyetso gishya, gihuza muburyo busanzwe hamwe na cache izunguruka yandika ibyanditswe kera kuruta gukura ubuziraherezo.
Ingaruka z'Ingamba
Umuvuduko n'igipimo
Ururimi rwakazi rushobora kugenda byihuse nta gutamba guhuzagurika.
Kugera no kugera
Yagura uburyo bwindimi nuburyo bwo gutumanaho.
Ibyemezo bisobanutse
Amakipe arashobora kumara umwanya munini murubanza mugihe automatike ikora gusubiramo.
Kazoza Kanyerera Idirishya Icyitonderwa
Ibishushanyo bya Hybrid ubu bihuza ibice bike byisi cyangwa byuzuye-byitondewe mubice byinshi byanyerera-idirishya, kuringaniza imikorere hamwe nukuri kure-ibitekerezo. Gemma 2 nabandi basimburana hafi yisi yose. Tegereza idirishya ryitondewe kugirango rihuze hamwe na leta-yumwanya wicyitegererezo, ibyerekezo byokwitabwaho, hamwe na KV-cache compression kugirango moderi yimipaka ikore miriyoni-token imiterere idafite ububiko bwo guhunga. Ihinduka inyubako isanzwe aho kuba exotic optimizme.
Gushyira mu bikorwa Isi
Mistral 7B ikoresha idirishya rya 4.096-ryerekana kunyerera hejuru yurwego rwayo kugirango ikore ibisobanuro birebire bihendutse kubakoresha GPUs.
Longformer ikoresha idirishya ryitondewe hiyongereyeho ibimenyetso bike kwisi kugirango ushyire hamwe kandi uvuge muri make impapuro nyinshi.
Gemma 2 ihinduranya kunyerera-idirishya ryibanze hamwe nisi-yita kumurongo kugirango iringanize umuvuduko hamwe nigihe kirekire cyo kwibuka.
Rolling-buffer KV cache mubufasha bunganira ibika gusa idirishya rya vuba ryibimenyetso, gufata ububiko mugihe cyibiganiro birebire.
Ingaruka & Kurinda
Ibintu bifatika bishobora kwinjiza bucece raporo, gushyigikira imigendekere, cyangwa ibisubizo byubushakashatsi.
Kwihuta byihuse birashobora gukora ibisubizo bidahuye mubisabwa bisa.
Ibyanditswe byumvikana birashobora kugaragara niba kugenzura kugenzura ari ntege.
Igishushanyo mbonera
Sobanura imiterere isohoka, amajwi, hamwe nubuziranenge mbere yo gutangira.
Ibisubizo byibanze hamwe nisoko yizewe igihe cyose ukuri kwingirakamaro.
Komeza kugenzura abantu kugenzura ibisubizo byinshi.
Kurikirana uburyo bwo kunanirwa no kongera imyitozo cyangwa akazi gahoraho.
Komeza Ubushakashatsi
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 Sliding Window 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
Ubuyobozi bukurikira
Itsinda-Ikibazo
Ibibazo bikunze kubazwa
What is Sliding Window Attention?
Kunyerera idirishya ryitondewe rigabanya buri kimenyetso cyo kwitabira gusa mubunini-bunini buturanye bwibimenyetso byegeranye aho kugirango bikurikirane. Ibi bigabanya ibiciro bya kwadratike yibitekerezo bisanzwe kumurongo, bigatuma imiterere-ndende-ndende ihendutse gukora.
Ni izihe nyungu nyamukuru zo kubara zo kunyerera idirishya ryitondewe hejuru yo kwiyitaho bisanzwe?
Mugabanye buri kimenyetso kumadirishya ihamye yabaturanyi W, igiciro cyiyongera nkuko N inshuro W (umurongo muri N) aho kuba N-kare.
Mu gishushanyo cya Mistral 7B, nigute amakuru ashobora kugenda kurenze idirishya rimwe 4.096?
Kimwe na CNN, buri cyiciro gisunika amakuru idirishya rimwe, bityo L ibice bitanga umwanya mwiza wikigereranyo cya L inshuro W.
Kuki kunyerera idirishya ryitondewe bigabanya kwibuka mugihe cyo kubyara inyandiko?
Kubera ko ikimenyetso cyitabira gusa idirishya ryacyo rya vuba, urufunguzo rushaje / agaciro ibyinjira birashobora gutabwa, akenshi binyuze mukuzunguruka.
Ni ubuhe buryo bwo guhitamo moderi nka Gemma 2 ikoresha kuruhande rwa Windows kunyerera kugirango ukomeze gutekereza kure?
Kuvanga rimwe na rimwe kwisi / byuzuye-byitondewe murwego rwibanze bikomeza guhuza intera ndende ihuza neza.
Kuburyo bwo kunyerera idirishya ryubunini W, ni uruhe rufunguzo rushobora kubaza kumwanya nitabira?
Mugihe cyimpamvu ikibazo kibona ubwacyo hamwe na W minus 1 ibimenyetso ako kanya mbere yacyo, ntabwo ibimenyetso bizaza.