Guhagarika-Ntibisanzwe na Kavukire Ntibisanzwe
Guhagarika-gake kandi kavukire gake reka reka abahindura bitabira gusa ibice byingenzi byurwego rurerure aho kuba ikimenyetso cyose, kugabanya ibiciro bya quadratic yo kwitabwaho bisanzwe.
Incamake
This is what makes efficient long-context models practical on real hardware.
Kwibira cyane
Kwiyitaho bisanzwe kugereranya buri kimenyetso nikindi kimenyetso, bityo igiciro kikura inshuro enye hamwe nuburebure bwikurikiranya, bigahinduka kubuza inyandiko ndende cyane. Kwitonda gake kugabanya buri kimenyetso kumurongo wabandi. Guhagarika-gake inzira igabanya urukurikirane mubice hanyuma ubare ibitekerezo gusa kubice byatoranijwe byombi, bishushanya neza kuri GPU tensor. Kavukire ya Sparse Attention (NSA), kuva muri DeepSeek, iragenda ikomeza: irashobora gutozwa impera-iherezo-iherezo kandi igahuza ibyuma, igahuza amashami atatu, guhunika-token-token compression, guhitamo neza-guhitamo neza byingenzi, hamwe nidirishya ryo kunyerera kurwego rwaho. Kuberako uburyo butandukanye bwigishwa mugihe cyo kwitegura aho guhinduka nyuma, NSA ibika neza mugihe itanga umuvuduko munini kumurongo muremure.
Ubushishozi
NSA itunganya urufunguzo nagaciro binyuze munzira eshatu zibangikanye, hanyuma ikabihuza namarembo yize. Kwiyunvira kwegeranya uduce twibimenyetso mubisobanuro byerekana; guhitamo amanota arahagarika kandi agumana gusa urutonde-rwo hejuru kugirango yitabweho byuzuye; idirishya ryinyerera ritwikiriye ibimenyetso. Guhagarika urwego rwibikorwa bihuza na GPU yibuka hamwe na tensor-core yinjiza, bityo kuzigama kwa FLOP kuzigama bisobanurwa muburyo bwihuse bwurukuta rwisaha mugihe cyamahugurwa no gufata umwanzuro, cyane cyane kubikorwa byo kwibuka.
Ingaruka z'Ingamba
Igiciro na bije
Ibyemezo byubwubatsi bitwara imikorere nigiciro cyimikorere kumyaka.
Ibyemezo bisobanutse
Ubuhanga bwa tekinike bufasha amakipe guhitamo umurongo ukwiye, ntabwo ari shyashya gusa.
Kugenzura ubuziranenge
Guhitamo neza bya injeniyeri bigabanya ibintu byizewe mubikorwa.
Ejo hazaza hahagaritswe-Ntibisanzwe na Kavukire Ntibisanzwe
Gutoza, ibyuma-bimenyesha sparsity birahinduka inzira ya miriyoni-token imiterere idaturika ikiguzi. Witege ko bidakunze kwitabwaho kugirango uhuze hamwe na kernel na yihuta, uhujwe n'umurongo-wo kwitaho hamwe n'ibitekerezo bya leta-umwanya, kandi bigakoreshwa mumipaka miremire-ndende hamwe n'ibitekerezo. Mugihe ibishushanyo bigenda byigwa kandi bigahinduka, icyitegererezo kizagenera ingengo yimari ijyanye nibibazo kuri buri kibazo, kandi ibipimo bizagenda bipima decoding yinjira kumurongo muremure, ntabwo ari ubuziranenge gusa.
Gushyira mu bikorwa Isi
Gukoresha icyitegererezo hejuru ya codebase yose cyangwa amasezerano maremare yemewe aho kwitabwaho byuzuye byananiza GPU yibuka.
NSA ya DeepSeek yihutisha kwitegura no gufata intera ndende mugihe ihuye cyangwa ikubita neza-ibitekerezo byuzuye.
Incamake yibitabo byuburebure witabira kwitabira incamake yo guhagarika wongeyeho ibice bifitanye isano.
Kwihutisha imiterere-ndende yo gufasha abafasha intambwe ya decoding intambwe ijyanye no kwibuka kugabanya buri kimenyetso kumurongo wo hejuru.
Ingaruka & Kurinda
Gutezimbere igipimo kimwe gishobora guhisha intege nke za sisitemu.
Ibikorwa Remezo no kubungabunga akenshi usanga bidahabwa agaciro.
Icyuho cyumutekano no kwitegereza birashobora kwiyongera uko sisitemu igenda igorana.
Igishushanyo mbonera
Sobanura ubukererwe, ubuziranenge, nigiciro cyibiciro mbere yo kubishyira mubikorwa.
Ibipimo byerekana umutwaro ufatika hamwe namakuru yimiterere.
Gukurikirana ibikoresho kubikosa, drift, ningaruka zabakoresha.
Tegura inzira yo gusubiza ibyabaye mbere yo gupima.
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 Block-Sparse and Native Sparse 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
Ntibisanzwe
Ibibazo bikunze kubazwa
What is Block-Sparse and Native Sparse Attention?
Guhagarika-gake kandi kavukire gake reka reka abahindura bitabira gusa ibice byingenzi byurwego rurerure aho kuba ikimenyetso cyose, kugabanya ibiciro bya quadratic yo kwitabwaho bisanzwe. Nibyo bituma moderi ikora neza-ndende igaragara mubikorwa byuma nyabyo.
Ni ukubera iki kwiyitaho bisanzwe bihenze kumurongo muremure?
Ikimenyetso cyose cyitabira ikindi kimenyetso, kubara rero no kwibuka umunzani hamwe na kare yuburebure bwikurikiranya.
Ni ikihe gitekerezo cyibanze cyo guhagarika ibitekerezo?
Urukurikirane rwacitsemo ibice kandi kwitabwaho kubarwa gusa kubice byatoranijwe byombi, nabyo bishushanya neza kuri GPU tensor.
Niki gituma Kavukire idasanzwe (NSA) itandukana nuburyo bwinshi bwambere?
NSA yiga uburyo bworoshye mugihe cyo kwitegura kandi igenewe guhuza ibyuma, bityo ikabika neza mugihe ikora byihuse.
Ni ayahe mashami atatu NSA ihuza urufunguzo n'indangagaciro?
NSA ihuza ikomatanya rito, guhitamo neza, hamwe nidirishya ryanyerera ukoresheje amarembo yize.
Ni ukubera iki NSA ikoresha ibikorwa byo guhagarika urwego aho gukoresha ibimenyetso-buke buke?
Guhagarika urwego rwo kugera kubikoresho bya GPU, kubwibyo kuzigama kwa FLOP bihinduka urukuta-isaha yihuta.