Toshe-Sparse da Hankali na ɗan ƙasa
Toshe-tabbataccen hankali da hankali na asali yana barin masu taswira su halarci mafi dacewa guntu-guntu na dogon jeri maimakon kowane alama, suna ɓata farashi mai ƙima na daidaitaccen kulawa.
Dubawa
Wannan shi ne abin da ke sa ingantattun samfuran dogon mahallin su zama masu amfani a kan kayan aiki na ainihi.
Zurfafa nutsewa
Daidaitaccen kulawar kai yana kwatanta kowane alama da kowace alama, don haka farashi yana girma da yawa tare da tsayin jeri, yana zama haramun ga dogon takardu. Hankali mara kyau yana taƙaita kowane alamar zuwa ɓangaren wasu. Hanyoyi masu toshewa suna rarraba jeri zuwa tubalan kuma suna ƙididdige hankali kawai don zaɓin toshe nau'i-nau'i, waɗanda ke yin taswira yadda ya kamata akan GPU tensor cores. Native Sparse Attention (NSA), daga DeepSeek, ya ci gaba: yana da horo na ƙarshe-zuwa-ƙarshe da daidaita kayan masarufi, haɗe rassa uku, damtse-ƙarfe-ƙasa, zaɓi mai kyau na mafi mahimmanci tubalan, da taga mai zamewa don mahallin gida. Saboda ana koyan tsarin ɓata lokaci a lokacin horo maimakon a rufe shi daga baya, NSA tana kiyaye daidaito yayin isar da babban saurin gudu akan dogayen jeri.
Fahimtar Fasaha
NSA tana aiwatar da maɓalli da ƙima ta hanyoyi guda uku masu kamanceceniya, sannan ta haɗa su da ƙofofin koyo. Matsi yana tara tubalan alamu zuwa taƙaitaccen wakilci; Zaɓin ya toshe ƙima kuma yana kiyaye manyan masu matsayi kawai don cikakkiyar kulawa; taga mai zamiya yana rufe alamun kusa. Ayyukan matakan toshewa suna daidaitawa tare da damar ƙwaƙwalwar ajiya na GPU da kayan aiki na tensor-core, don haka tanadin FLOP na ka'idar yana fassara zuwa ainihin saurin agogo na bango yayin duka horo da ƙima, musamman don matakin ƙaddamar da ƙwaƙwalwar ajiya.
Dabarun Tasiri
Kudin da kasafin kuɗi
Hukunce-hukuncen gine-gine suna haifar da aiki da tsadar aiki na shekaru.
Shawarwari masu haske
Ilimin fasaha yana taimaka wa ƙungiyoyi su zaɓi tari mai kyau, ba kawai sabon abu ba.
Kula da inganci
Zaɓuɓɓukan injiniya mafi kyau suna rage abin dogaro a cikin samarwa.
Makomar Block-Sparse da Hankali na 'Yan Ƙasa
Mai horarwa, rashin sanin kayan masarufi yana zama hanyar zuwa mahallin alamar miliyoyin ba tare da fashe farashi ba. Yi tsammanin kulawar da ba ta dace ba za a tsara shi tare da kernels da masu haɓakawa, hade da hankali-hankali da ra'ayoyin sararin samaniya, kuma a karbe su a cikin yanayi mai tsayi da ƙirar tunani. Yayin da alamu ke zama abin koyo da kuzari, ƙila za su keɓance kasafin kuɗi daidai gwargwado ta kowace tambaya, kuma maƙasudi za su ƙara auna ƙaddamar da kayan aiki akan dogayen jeri, ba kawai ɗanyen inganci ba.
Aiwatar da Gaskiyar Duniya
Gudanar da samfuri akan gabaɗayan codebase ko kwangilar doka mai tsawo inda cikakkiyar kulawa zata ƙare ƙwaƙwalwar GPU.
DeepSeek's NSA yana haɓaka duka pretraining da kuma dogon tunani yayin daidaitawa ko bugun cikakkiyar kulawa.
Taƙaita takaddun tsawon littafi ta hanyar halartar taƙaitaccen taƙaitaccen bayanin toshe tare da sassan da suka dace na cikin gida.
Ƙaddamar da mataimakan taɗi na dogon lokaci waɗanda matakin yanke hukunci yana da alaƙa da ƙwaƙwalwar ajiya ta iyakance kowane alamar zuwa manyan tubalan.
Hatsari & Tsare-tsare
Haɓaka ma'auni ɗaya na iya ɓoye manyan raunin tsarin.
Sau da yawa ana raina kayan more rayuwa da kuma kuɗin kulawa.
Tsaro da gibin lura na iya girma yayin da tsarin ke ƙara haɓaka.
Taswirar Hanya
Ƙayyade latency, inganci, da maƙasudin farashi kafin aiwatarwa.
Alamar ma'auni a ƙarƙashin ainihin kaya da yanayin bayanai.
Kula da kayan aiki don kurakurai, ɗigo, da tasirin mai amfani.
Shirya bijirowa da hanyoyin mayar da martani kafin sikeli.
Ci gaba da Bincike
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
Jagora na gaba
Tsare-tsare Hankali
Tambayoyin da ake yawan yi
Menene Block-Sparse da Native Sparse Attention?
Toshe-tabbataccen hankali da hankali na asali yana barin masu taswira su halarci mafi dacewa guntu-guntu na dogon jeri maimakon kowane alama, suna ɓata farashi mai ƙima na daidaitaccen kulawa. Wannan shine abin da ke sa ingantattun samfuran yanayi na dogon lokaci a aikace akan kayan aiki na gaske.
Me yasa daidaitaccen hankalin kai yana da tsada don dogon jerin?
Kowane alamar yana halartar kowace alama, don haka ƙididdige da sikelin ƙwaƙwalwar ajiya tare da murabba'in tsayin jeri.
Menene ainihin ra'ayin toshe-raguwar hankali?
An raba jerin zuwa tubalan kuma ana ƙididdige hankali ne kawai don zaɓin toshe nau'i-nau'i, wanda kuma ya yi taswira da kyau akan jigon GPU tensor.
Me ya sa Hankalin Ƙarƙashin Ƙasa (NSA) ya bambanta da yawancin hanyoyin da ba su da yawa a baya?
NSA tana koyon yanayin rashin ƙarfi yayin horo kuma an ƙirƙira shi don daidaitawa da kayan aiki, don haka yana kiyaye daidaito yayin aiki da sauri.
Wadanne rassa uku ne NSA ta hada don makullinta da kimarta?
NSA tana haɗa ƙaƙƙarfan matsawar alamar, zaɓin shinge mai kyau, da taga zamiya ta gida ta amfani da ƙofofin ilmantarwa.
Me yasa NSA ke amfani da ayyukan toshewa maimakon rashin daidaituwar matakin matakin saɓani?
Hanyoyin isa ga matakin toshewa sun dace da kayan aikin GPU, don haka tanadin FLOP na ka'idar ya zama ainihin saurin agogon bango.