Medusa boppu dekodaas
Medusa xeetu dekodaasu speculatif la buy boole ay 'bopp' yuy wax luy waaja am ci benn xeetu làkk suko defee mu mëna xayma ay jeton yu bari ci ëlëg benn yoon.
Résumé
By verifying these guesses in a single forward pass, it speeds up text generation roughly 2-3x without changing the model's output distribution.
Plongeur bu xóot
Modèlu làkk yi ñuy jëfandikoo dañuy defar benn token ci paas bu nekk, te loolu dafay yeex ndax jéego bu nekk dafa wara xaar jéego bu njëkk bi. Medusa dafa yokk boppu feed-forward yu woyof ci kaw modelu base bu liw; bopp bu nekk dafay wax luy waaja am ci position yu néew ci kanam (bopp 1 dafay wax luy waaja am ci token bi ci topp, bopp 2 ci token bi ci topp, ak ñoom seen). Xalaat yooyu dañuy nuru garabu wéyal kàndidaa yi. Ginaaw loolu model bi dafay xool garab gi yépp ci benn yoon, jëfandikoo mask 'tree attention', nangu prefix bi gëna gudd bi méngoo ak li model bi wara defar. Ndax xool dafay jëfandikoo xeetu njëkk ba, Medusa amul benn perte: bind bi ñu nangu mooy li decodage bu bëgg xaalis wala sampled bi di génne, ñu defar ko ci jéego yu néew yu toppalante.
Gis-gis xarala
Boppu Medusa bu nekk ab MLP residuel bu ndaw la buy màndargaal tolluwaayu nëbbu bu mujj bi ci modelu base bi ci séddale ci kaw ay token ci offset k. Candidat yi ci bopp yi dañu leen raññe ci garab, ak masku bàyyi xel buñ defar ngir may modelu base bi mu poñ car bu nekk ci benn yoon ci benn yoon. Benn pexem nangu buñ miin mooy tànn ban jeton lañu wara tëye, ngir gaaraati ni resultaa bi méngoo ak sampling modelu base bi, suko defee kalite bi di gëna am solo bi jéego yu toppalante yi di wàññeeku.
njeextalu pexe
Gaawaay ak yaatuwaay
Liggéeyukaay yi ci làkk yi mën nañu gëna gaaw te duñu yàq deggoo gi.
Dugg ak yegg
Dafay yaatal jëfandikoo gi ci làkk yi ak ci anam yi ñuy jokkoo.
dogal yu gëna leer
Ekip yi mën nañu gëna yàgg ci àtte ci jamono ji otomatisation di liggéey ci baamtu.
Ëlëgu boppu dekodaayu Medusa
Decodage speculatif mingi nekk standard ci stack inference yi, ak jegewaale yu nekk ci seen bopp yu melni Medusa, yu moytu soxla modelu projet bu wuute, dañuy xëcc nit ñi ndax ñoo gëna yomba jëfandikoo. Liggéeyu ëlëg dafay boole boppu yu nuroo ak Medusa ak xam-xam bu nuroo ak EAGLE, tabax garab gu gëna baax, ak xool bu baax ci hardware. Xaarandi lëkkaloo bu gëna dëgër ci biir kaadar yiy liggéey, tuning otomatik ci jëmmu garab ci liggéey bu nekk, ak boole ak KV-cache compression suko defee latency wàññeeku te du am yeneen GPU wala ñàkka am kalite.
Doxal ci àdduna dëgg
Dagg tontu chatbot latency ci nangu jetons yu bari yuñ firndeel ci pass forward bu nekk
Gaawaale ndimmbalu matt kode fu ñu mëna seentu token yu yomb di xalaat
Wàññi njëgu inference ngir API LLM yu bari trafik te duñu dugal benn xeetu projet bu wuute
Gaawaale defar mbind yu gudd yu melni ay resumé boole ci tëye génne gi mu méngoo ak dekodaas buñ miin
Risk yi ak balustrade yi
Lépp lu jaarul yoon mën na dugg ci rapoor yi, jàppale ci liggéey bi, wala ci njariñu gëstu bi.
Sensibilite bu gaaw mën na jur njariñ yu wuute ci laajte yu noonu mel.
Done yu am solo mën nañu feeñ sudee seytu jëfandikoo gi néew doole.
Roadmap ngir samp gi
Mandargal formaa génne gi, melokaan bi, ak standard kalite yi laata ngay dugal ko.
Tontu yu am solo ak balluwaay yu wóor saa yu dëggu bi di am solo.
Fexeel am barabu xool nit ñi ngir am njariñ yu am solo.
Toppal anami gacce yi ak di faral di tàggataat ay laaj wala def-liggéey.
Weyal di banneexu
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 Medusa Decoding Heads 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
Gis bi ci topp
Baamtu Penalty ak Dekodage Kontrol
Laaj yi ñuy faral di laaj
What is Medusa Decoding Heads?
Medusa xeetu dekodaasu speculatif la buy boole ay 'bopp' yuy wax luy waaja am ci benn xeetu làkk suko defee mu mëna xayma ay jeton yu bari ci ëlëg benn yoon. Soo xoolee bu baax xalaat yooyu ci benn jéego, dafay gaaw defar mbind lu tollu ci 2-3x te du soppi distribution bi ci model bi.
Lan la boppu Medusa yu bari yi di wax?
Boppu Medusa bu nekk dafay wax luy waaja am ci position yu bari ci ëlëg, kon mën nañu proposer ay token yu bari benn yoon.
Lan moo tax ñu jàpp ni Medusa 'amul benn ñàkk' buñu ko méngale ak dekodaas buñ miin?
Modèle de base dafay saytu jetons candidat yi, nangu rek ñooñu mu wara defar, ba noppi baña yàq distribution biy génn.
Ci yan anam lañuy amalee wéyal kàndidaa yu Medusa ngir saytu?
Candidat yi dañuy forme garab, masku bàyyi xel ci garab gi dafay may modelu base bi mu xool bépp car ci benn jéego bu jëm kanam.
Lan mooy njariñu Medusa ci dekodaasu spekulatif bi?
Medusa dafay takk bopp yu woyof yi ci model bi fi nekk, kon jarul ñu tàggat ak liggéey ci reso bu wuute.
Lu tollu ci ban gaawaay la Medusa faral di xamle?
Medusa dafay wàññi lu tollu ci 2 ba 3 yoon ci latency bi, lépp di aju ci ni liggéey bi di doxee.