Ƙwarewar Hasashen da EAGLE
Ƙididdigar ƙididdiga na haɓaka babban ƙirar ƙirar harshe ta hanyar barin ƙaramin ƙirar ƙirar ƙira da alamu da yawa a gaba, waɗanda babban samfurin sannan ya tabbatar da wucewa ɗaya.
Dubawa
EAGLE is a state-of-the-art version that drafts at the feature level rather than the token level, delivering 2-4x speedups with zero loss in output quality.
Zurfafa nutsewa
Ƙarfin LLM na al'ada yana da juzu'i: samfurin yana samar da alama ɗaya, yana ciyar da shi, kuma yana maimaitawa, don haka kowace alamar tana buƙatar cikakken gaba ta hanyar biliyoyin sigogi. Ƙididdigar ƙira ta karya wannan ƙulli. Mai ƙira mai arha yana ba da ɓangarorin alamun ɗan takara, kuma ƙirar manufa mai tsada tana tabbatar da duka a cikin layi ɗaya na layi ɗaya, yana karɓar madaidaicin prefix mafi tsayi. EAGLE (Extrapolation Algorithm for Greater Language-Model Efficiency) yana inganta hanyoyin da suka gabata ta hanyar zayyana sararin fasalin fasalin da ke ɓoye da kuma ciyar da abin da ya gabata na gaskiya na token don rage rashin tabbas. EAGLE-2 yana ƙara bishiyar daftarin aiki mai ƙarfi, kuma EAGLE-3 yana sauke ƙayyadaddun yanayin-hasashen don auna mafi kyau. Mahimmanci, tabbatarwa yana ba da tabbacin abin da aka fitar ya yi daidai da abin da ƙirar da aka yi niyya za ta samar ita kaɗai.
Fahimtar Fasaha
EAGLE yana horar da ƙaramin shugaban kai-da-kai wanda yayi hasashen fasalin tsarin ɓoyayyiyar jihar gaba na gaba, sannan ya sake yin amfani da kan LM na kansa wanda aka yi niyya don juya fasali zuwa ƴan takara masu alama. Ta hanyar daidaita tsarin jujjuyawar alama tare da abubuwan da suka gabata, yana yanke shubuhar da ta addabi tsara fasalin-kawai. Ana tabbatar da bishiyar 'yan takara lokaci guda; Rarraba samfurin da aka yi niyya ana kiyaye shi daidai saboda karɓaɓɓun alamun dole ne su dace da abin da aka ƙirƙira ko zaɓin argmax, sa saurin gudu ya zama mara asara.
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 Ƙwarewar Hasashen tare da EAGLE
Ƙididdigar ƙididdigewa yana zama tsoffin abubuwan more rayuwa a cikin hidimar tarawa kamar vLLM da TensorRT-LLM. Yi tsammanin haɗin kai mai ƙarfi tare da batching da raba-cache KV, ƙirar ƙira waɗanda ba su buƙatar mai ƙira daban-daban, da ƙirar haɗin gwiwar kayan aiki waɗanda ke ɗaukar tabbatarwa iri ɗaya. Ana ƙaddamar da tsara fasalin fasalin EAGLE zuwa nau'ikan abubuwa da yawa da tunani, inda dogayen sarƙoƙi na tunani ke yin tsadar kowane alama musamman mai raɗaɗi, da kuma ba da shawarar kan na'urar inda latency ya fi dacewa.
Aiwatar da Gaskiyar Duniya
Yanke latency a cikin mataimakan hira don haka martani yana gudana 2-3x cikin sauri ba tare da canza amsoshin samfurin ba.
Rage farashin sabis na GPU don masu samar da API mai girma ta hanyar samar da ƙarin alamu a kowane wucewar gaba
Hanzarta dogayen tsarin tunani na tunani inda ake samar da dubban alamu a kowace tambaya
Ƙaddamar da kayan aikin kammala lambar inda za'a iya iya faɗi, jerin alamu masu maimaitawa suna haifar da ƙimar yarda da ƙima.
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 Speculative Decoding with EAGLE 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
Ƙididdigar Hasashen
Tambayoyin da ake yawan yi
What is Speculative Decoding with EAGLE?
Ƙididdigar ƙididdiga na haɓaka babban ƙirar ƙirar harshe ta hanyar barin ƙaramin ƙirar ƙirar ƙira da alamu da yawa a gaba, waɗanda babban samfurin sannan ya tabbatar da wucewa ɗaya. EAGLE sigar zamani ce ta zamani wacce ke zayyana a matakin fasalin maimakon matakin alamar, yana ba da saurin gudu na 2-4x tare da asarar sifili a ingancin fitarwa.
Menene ainihin ra'ayin da ke bayan ƙwaƙƙwaran ƙima?
Ƙaramin mai ƙira yana hasashen alamu da yawa a gaba, kuma babban samfurin manufa yana tabbatar da su tare, yana karɓar madaidaicin madaidaicin mafi tsayi.
Ta yaya EAGLE ya bambanta da hanyoyin tantancewa na baya?
EAGLE yana tsinkayar ɓoyayyun ƙirar ƙirar da aka yi niyya kuma yana sake amfani da kan LM ɗin sa, wanda ke samar da ingantattun zane fiye da hanyoyin alama kawai.
Me yasa ake ɗaukar zazzage ƙididdiga a matsayin 'marasa asara'?
Saboda ƙirar da aka yi niyya tana duba kowace alamar da aka tsara ta hanyar rarraba ta, abin da aka karɓa shine ainihin abin da manufa za ta haifar shi kaɗai.
Wace matsala a cikin tsara fasalin EAGLE zai taimaka wajen magance abin da ya gabata?
Ƙaddamarwa akan ainihin alamar da ta gabata tana rage shakku na tsinkayar siffa mai ɓoye na gaba, inganta ingantaccen daftarin aiki.
Menene EAGLE-2 ya ƙara akan ainihin EAGLE?
EAGLE-2 ya gabatar da bishiyar daftarin yanayi mai sane, yana faɗaɗa rassa masu ban sha'awa don haɓaka ƙimar karɓa.