Overview
Kutepfenyura uye kufanotaura kwezviratidzo zvakawanda kumhanyisa chizvarwa chemhando yemutauro nekufungidzira akati wandei emangwana tokeni kamwechete uye nekuaona nekupfuura kumwe chete, pane kugadzira chiratidzo chimwe panguva. Ivo vanocheka latency vasina kushandura zvinyorwa izvo modhi ingadai yakanyora.
Kufungidzira Kutenderera uye Multi-Token Prediction idhizaini yekuvaka inobata mhando yemhando, mutengo wezvivakwa, latency, uye kuvimbika pachiyero.
Deep Dive
Yakajairika autoregressive decoding inononoka nekuti chiratidzo chega chega chinoda kupfuudza kumberi uye ma tokeni anogadzirwa akanyatso mumwe mushure memumwe, achisiya GPU isingashandiswe. Kudhirodha kwekufungidzira kunogadzirisa izvi nedhizaini rakachipa iro rinokurudzira chunk yematokeni evamiriri, iyo iyo hombe yaivavarirwa modhi yobva yaona zvakafanana; chero prefix inofanana neicho chinangwa chingave chaburitsa chinogamuchirwa mahara, uye yekutanga kusawirirana inogadziriswa. Kutepfenyura kutenderera uye Medusa-maitiro akawanda-tokeni kufanotaura peta dhizaini mumodhi pachayo: yakawedzera kureruka misoro yekufungidzira (kana rwizi rwekufungidzira tokens) rega modhi imwe inyore uye simbisa, ichinzvenga yakaparadzana dhizaini modhi. Nekuti verification ndeyechokwadi, kugovera kwakabuda kwakafanana neyakajairwa decoding, iwe unongowana 2 kusvika ku3 kakapetwa mashoma anotevedzana matanho.
Technical Insight
Chinokosha ndechekuti transformer inogona kukwevera zvinzvimbo mune imwe yekumberi kupfuura zvakachipa seimwe, sezvo iri ndangariro-bandwidth yakasungwa, kwete compute yakasungwa, panguva yekunyora. Misoro yakawanda yekufembera inoburitsa ma tokens ezvinzvimbo zvinoverengeka zvinotevera; muti kana kutevedzana kwevanoda kukwikwidza kunosimbiswa pamwe chete, uye kugamuchirwa kunoshandisa sampling yekuramba (kana makaro kuenzanisa) kuitira kuti zviratidzo zvinogamuchirwa zviteedzere kugovera kwakanangana. Kureba kunogamuchirwa panhanho kunosarudza kukurumidza.
Kubata Kufungidzira Kutenderera uye Multi-Token Prediction
Kutepfenyura uye kufanotaura kwezviratidzo zvakawanda kumhanyisa chizvarwa chemhando yemutauro nekufungidzira akati wandei emangwana tokeni kamwechete uye nekuaona nekupfuura kumwe chete, pane kugadzira chiratidzo chimwe panguva. Ivo vanocheka latency vasina kushandura zvinyorwa izvo modhi ingadai yakanyora. Kufungidzira Kutenderera uye Multi-Token Prediction idhizaini yekuvaka inobata mhando yemhando, mutengo wezvivakwa, latency, uye kuvimbika pachiyero. Kuti uvake kunzwisisa kwakadzama, tora Kufungidzira Kutenderera uye Multi-Token Prediction semuenzaniso wekushandisa, kwete chinhu chimwe chete: tsanangura zvaunoda mhedzisiro, kujekesa fungidziro, uye patsanura zvinogona kuitwa nehurongwa hwakavimbika kubva kune izvo zvichiri kuda kutonga kwenyanzvi.
Mukuita, zvikwata zvakasimba zvinoshandisa Speculative Streaming uye Multi-Token Prediction inokwidziridza zvivakwa, data, uye sarudzo dzezvivakwa zvinopesana nekuvimbika uye mutengo. Ivo vanonyora zvakajeka maitiro ebudiriro, bvunzo vachipokana ne data rechokwadi uye mafambiro ebasa, uye iterate zvichibva pane zvakacherechedzwa maitiro ekutadza kwete kuhwina-nguva imwe chete yebhenji. Apa ndipo apo kunzwisisa kwe theoretical kunoshanduka kuve kugona kwakasimba pane chigadzirwa, mutemo, uye mashandiro.
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore. Panguva imwecheteyo, Kukwirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba. Nzira yakatsiga ndeyekubatanidza kukurumidza kuyedza nekutonga: mhanyisa vatyairi vendege, tora humbowo, buritsa matanda esarudzo, uye urambe uchivandudza chengetedzo semaitiro emuenzaniso, zvinotarisirwa nemushandisi, uye zvinodikanwa zvekutonga.
Strategic Impact
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore.
Zvisarudzo zvezvivakwa zvinotyaira kuita uye mutengo wekushandisa kwemakore. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete.
Dzidzo yehunyanzvi inobatsira zvikwata kusarudza murwi wakakodzera, kwete iwo mutsva chete. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira.
Sarudzo dzeinjiniya dziri nani dzinoderedza zviitiko zvekuvimbika mukugadzira. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Real-World Implementation
Kucheka mhinduro latency yemubatsiri wekutaura ne2 kusvika 3x uchishandisa Medusa-maitiro ekuwedzera ekufanotaura misoro.
Kuwedzera-yekufungidzira decoding kune inference server saka hapana yakaparadzana dhizaini modhi inoda kutambirwa
Kumhanyisa kupedzisa kodhi apo refu, inofanotaurwa tokeni inotambirwa muzvikamu zvakakura
Kuderedza mutengo weGPU pachikumbiro nekuburitsa mamwe ma tokens kubva kune yega yega-inosungwa yekupfuura pass
Maitiro Ekuita
Yekufungidzira Kutenderera uye Multi-Token Prediction mukuita
Kucheka mhinduro latency yemubatsiri wekutaura ne2 kusvika 3x uchishandisa Medusa-maitiro ekuwedzera ekufanotaura misoro.
Kucheka kunonoka kwekupindura kwemubatsiri wekutaura ne2 kusvika ku3x uchishandisa Medusa-maitiro ekuwedzera ekufanotaura misoro Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa nemitengo yekukanganisa nekufamba kwenguva.
Yekufungidzira Kutenderera uye Multi-Token Prediction mukuita
Kuwedzera-yekufungidzira decoding kune inference server saka hapana yakaparadzana dhizaini modhi inoda kutambirwa.
Kuwedzera yekuzvifungira decoding kune inference server saka hapana yakapatsanurwa dhizaini modhi inoda kutambirwa Matimu anowanzo kuwana zvirinani zvibodzwa kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa nemitengo yekukanganisa nekufamba kwenguva.
Yekufungidzira Kutenderera uye Multi-Token Prediction mukuita
Kumhanyisa kupedzisa kodhi apo refu, inofanotaurwa tokeni inotambirwa muzvikamu zvakakura.
Kumhanyisa kupedzwa kwekodhi uko marefu, anofungidzira tokeni anomhanya anogamuchirwa mumachunks makuru Matimu anowanzo kuwana mibairo iri nani kana achinge atsanangudza emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa nemitengo yekukanganisa nekufamba kwenguva.
Yekufungidzira Kutenderera uye Multi-Token Prediction mukuita
Kuderedza mutengo weGPU pachikumbiro nekuburitsa mamwe ma tokens kubva kune yega yega-inosungwa yekupfuura pass.
Kuderedza mutengo weGPU pachikumbiro nekubvisa mamwe ma tokens kubva kune yega yega-inosungwa-mberi yekupfuura Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Njodzi & Guardrails
Kugadzirisa imwe bhenji kunogona kuvanza yakafara system kushaya simba.
Infrastructure uye mari yekugadzirisa inowanzotarisirwa pasi.
Chengetedzo uye kucherechedzwa mapundu anogona kukura sezvo masisitimu anowedzera kuoma.
Implementation Roadmap
Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa.
Tsanangura latency, mhando, uye mutengo zvinangwa usati waitwa. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Benchmark pasi pechokwadi mutoro uye data mamiriro.
Benchmark pasi pechokwadi mutoro uye data mamiriro. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro.
Chishandiso chekutarisa zvikanganiso, kudonha, uye mushandisi maitiro. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera.
Gadzirira nzira dzekudzosera kumashure uye dzezviitiko usati wawedzera. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.