Multi-Umutwe Utinze Kwitonda
Multi-Head Latent Attention (Umudepite) nuburyo bwo kwitondera, bwatangijwe muri DeepSeek-V2, bugabanya kwibuka-ushonje urufunguzo-agaciro cache mukantu gato gasangiwe rwihishwa.
Incamake
It lets large language models run with far less GPU memory while keeping quality close to standard attention.
Kwibira cyane
Iyo transformateur itanga inyandiko, ibika urufunguzo nagaciro ka vector kuri buri kimenyetso cyashize muri 'KV cache.' Iyo cache ikura hamwe nuburebure bwimiterere kandi yiganjemo gukoresha kwibuka mugihe cyo gufata umwanzuro. Umudepite asimbuye byinshi byuzuye-urufunguzo / agaciro indangagaciro hamwe numurongo umwe wo hasi urwego rwihishwa rwerekanwa kuri buri kimenyetso, hanyuma imishinga ihita isubira mumfunguzo zumutwe hamwe nagaciro kuguruka. Kuberako ibice byihishe byihishe gusa, DeepSeek-V2 yatangaje ko igabanya ububiko bwa KV-cache hejuru ya 90% ugereranije nibisanzwe-imitwe yitabwaho, bigatuma ibice birebire hamwe nubunini bunini. Icy'ingenzi, matrices yo hejuru-ishobora gutegekwa mubindi bipimo, bityo rero Umudepite agera kuri compression hamwe nigihombo gito cyangwa ntagereranywa muburyo bwiza bwo kwerekana.
Ubushishozi
Umudepite akora urwego rwo hasi rwo guhuzagurika: buri kimenyetso cyihishe leta iteganijwe kugeza kumurongo muto wihishe, kandi matrices yo hejuru-projection yongeye kubaka urufunguzo numutwe. Amayeri yubwenge ni 'gukurura' uburemere buringaniye mubibazo nibisohoka, bityo icyitegererezo nticyigera kigaragaza urufunguzo rwuzuye / indangagaciro mugihe cyo gufata umwanzuro. Umwanya wo guhinduranya ibintu bifashishwa hamwe ninzira yurufunguzo, kubera ko kuzenguruka bidashobora kwakirwa kimwe, kubika amakuru yumwanya.
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 ka Multi-Umutwe Utinze Kwitonda
Umudepite yafashije gukora DeepSeek-V2 na V3 mu bukungu kugira ngo akorere ku rugero, kandi tekinike iragenda ikwirakwira mu gihe amakipe yirukanka bihendutse. Tegereza imiterere-yinteko ishinga amategeko yihishe kugirango ihuze hamwe gake-Uruvange-rw-Impuguke, cashe zingana, hamwe na decoding yibishushanyo mbonera bizaza. Abashakashatsi kandi barimo gushakisha uburyo urugero rwihishwa rushobora kugabanuka mbere yo kugabanuka kwiza, kandi niba igitekerezo kimwe cyo hasi gishobora guhagarika ibitekerezo mugihe cyamahugurwa, ntabwo ari umwanzuro gusa.
Gushyira mu bikorwa Isi
Gukora DeepSeek-V2 / V3 moderi yo kuganira hamwe na GPU yibuke yibirenge byibisobanuro kuri buri cyifuzo
Gukoresha inyandiko ndende ikibazo gisubiza aho cache nini ya KV ubundi yarangiza VRAM
Kongera ingano yicyiciro cya GPU ihamye kuko buri ruhererekane rubika gusa agace gato kihishe
Gushoboza imirongo miremire ya Windows kubikoresho byibicuruzwa kubisubiramo-byongerewe abafasha
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 Multi-Head Latent 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
Ibibazo byinshi
Ibibazo bikunze kubazwa
What is Multi-Head Latent Attention?
Multi-Head Latent Attention (Umudepite) nuburyo bwo kwitondera, bwatangijwe muri DeepSeek-V2, bugabanya kwibuka-ushonje urufunguzo-agaciro cache mukantu gato gasangiwe rwihishwa. Kureka ururimi runini rwimikorere ikora hamwe na GPU yibuke mugihe ukomeza ubuziranenge hafi yo kwitabwaho bisanzwe.
Ni ikihe kibazo cyibanze Multi-Head Latent Attention yagenewe kugabanya?
Umudepite yibasiye Kache cache, ikura hamwe nuburebure bwimiterere kandi ikagenga kwibuka mugihe cyo gutanga inyandiko.
Nigute Umudepite agabanya cache ya KV?
Umudepite yihishe icyerekezo kimwe cyihishwa kuri buri kimenyetso kandi yubaka urufunguzo nindangagaciro biva muri up-projection.
Niyihe moderi yabanje kwerekana Multi-Head Yatinze Kwitonda?
Umudepite yatangijwe na DeepSeek muburyo bwa DeepSeek-V2 hanyuma ajyanwa muri DeepSeek-V3.
Ni ukubera iki Umudepite akeneye inzira 'decoupled' itandukanye yo guhinduranya imyanya?
Guhinduranya kuzenguruka ntibishobora kwinjizwa mubindi bipimo byuburemere, bityo rero Umudepite agumana ikintu gito cyibanze cyibanze kugirango atware amakuru yumwanya.
Hafi ya KV-cache yo kugabanya ububiko bwa DeepSeek-V2 bwatanzwe na Depite hamwe nibisanzwe imitwe myinshi?
DeepSeek-V2 yatangaje ko igabanya ububiko bwa KV-cache hejuru ya 90%, bigatuma ibice birebire hamwe nitsinda rinini.