HejuruUbuyobozi bukurikira
Gusubiramo Ibihano no Kugenzura Igenzura
Ururimi AI
Ururimi AI
Medusa nuburyo bwo gutondeka-gushushanya uburyo butandukanye bwo guhanura 'imitwe' kurugero rwururimi kugirango rushobore gukeka ibimenyetso byinshi bizaza icyarimwe.
Umva iki gitaboAI-byavuzwe · ~2 min
By verifying these guesses in a single forward pass, it speeds up text generation roughly 2-3x without changing the model's output distribution.
Normal language models generate one token per forward pass, which is slow because each step must wait for the previous one. Medusa adds lightweight feed-forward heads on top of the frozen base model; each head predicts a token a few positions ahead (head 1 predicts the next token, head 2 the token after, and so on). These predictions form a tree of candidate continuations. The full model then verifies the whole tree in one pass using a 'tree attention' mask, accepting the longest prefix that matches what the model would have produced anyway. For Medusa-1, the backbone stays frozen and original-model rejection sampling can preserve the base model's output distribution, enabling lossless inference acceleration. Medusa-2 jointly trains the heads and backbone with a special recipe that preserves the backbone's next-token capability while improving head prediction. The paper also studies typical acceptance, which can increase acceptance and maintain similar generation quality but does not make the same exact distribution-matching guarantee.
Gukata chatbot igisubizo cyatinze mukwemera ibimenyetso byinshi byagenzuwe kuri pass imbere
Kwihutisha code-kurangiza abafasha aho guhanura ibimenyetso bikurikirana byoroshye kubitekerezaho
Kugabanya igiciro cyimyanzuro yimodoka nyinshi LLM APIs udakoresheje umushinga wihariye
Kwihutisha imiterere miremire yinyandiko nkincamake mugihe ibisohoka bisa na decoding isanzwe
Gukoresha decoding yibintu bigenda bihinduka mubisanzwe byerekana umusaruro, kandi uburyo bwonyine bwonyine nka Medusa, bwirinda gukenera umushinga wihariye, burashimishije kuko bworoshye kubwohereza. Akazi kazoza gahuza imitwe-yuburyo bwa Medusa hamwe nuburyo bwa EAGLE buranga guhanura, kubaka ibiti neza, hamwe no kugenzura ibyuma. Tegereza kwishyira hamwe muburyo bwo gutanga serivisi, guhuza byikora kumiterere yibiti kumurimo wakazi, no guhuza hamwe na KV-cache compression kugirango ubukererwe bugabanuke nta GPU yongeyeho cyangwa gutakaza ubuziranenge.
Ibaruwa y'amakuru y'ubuntu
Inkuru eshatu zizewe za AI buri gitondo cy'iminsi y'akazi, zanditse mu rurimi rworoshye. Ubuntu iteka ryose, nta kwamamaza.
Imeri imwe buri munsi y'akazi. Kwiyimura mu buryo bumwe. Ntitwigera tugurisha cyangwa ngo dusangire aderesi yawe.
Isuzume
Ibisubizo ako kanya kuri buri kibazo, n'icyemezo gishobora gusangirwa gifite ID yagenzurwa umaze gutsinda isomo.
Fasha uburezi bwa AI bw'ubuntu. AI Understanding ni umuryango utegamiye kuri leta wa 501(c)(3) — nta kwamamaza, nta kwishyura, iteka ryose. Tanga impano
Medusa nuburyo bwo gutondeka-gushushanya uburyo butandukanye bwo guhanura 'imitwe' kurugero rwururimi kugirango rushobore gukeka ibimenyetso byinshi bizaza icyarimwe. Mugenzura ibi bitekerezo mumurongo umwe wimbere, byihutisha ibisekuruza byanditse hafi 2-3x idahinduye icyitegererezo cyo gukwirakwiza.
Buri mutwe wa Medusa uhanura ikimenyetso cyimyanya myinshi mugihe kizaza, bityo ibimenyetso byinshi birashobora gutangwa icyarimwe.
The Medusa paper identifies Medusa-1, which freezes the backbone, as enabling lossless acceleration; exact rejection sampling verifies accepted candidates against the base model distribution.
Abakandida bagize igiti, kandi igiti cyo kwitondera igiti kireka icyitegererezo kigenzura amashami yose muri pass imwe.
Medusa ifata imitwe yoroheje kuri moderi iriho, ntabwo rero bikenewe guhugura no gutanga umushinga wihariye.
Medusa muri rusange igera hafi ya 2-3x igabanya ubukererwe bwibisekuru bitewe numurimo mukazi.
Komeza wige
Abandi bayobozi batoranijwe kuriyi ngingo
HejuruUbuyobozi bukurikira
Gusubiramo Ibihano no Kugenzura Igenzura
Ururimi AI