InoteveraGaidhi rinotevera
Kududzira Kwokufungidzira
Tekinoroji
Nhungamiro yehunyanzvi
Kudhirodha kwekufungidzira kunomhanyisa modhi yemutauro wakakura nekuita kuti modhi diki diki ifungidzire ma tokeni akati wandei kumberi, ayo modhi hombe yobva yaona mune imwe pass.
Teerera kugwaro iriAI-yakarondedzerwa · ~2 min
EAGLE imhando-ye-iyo-iyo vhezheni inodhirowa pachikamu chechikamu kwete chechiratidzo, kuendesa 2-4x kumhanya nekurasikirwa zero mumhando yekubuda.
Chizvarwa cheLLM chinowanzoita autoregressive: modhi inogadzira chiratidzo chimwe, ichidyisa zvakare, uye inodzokorora, saka chiratidzo chega chega chinoda kumberi kuzere kuburikidza nemabhiriyoni emaparamita. Kudhirodha kwekufungidzira kunotyora dambudziko iri. Dhizaini yakachipa inokurudzira chunk yematokeni evamiriri, uye iyo inodhura yemhando yemhando inosimbisa ese mune imwechete yakafanana pass, achigamuchira iyo yakareba chaiyo prefix. EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency) inovandudza panzira dzekare nekunyora munzvimbo yakavanzika yemuenzaniso nzvimbo uye kudzoreredza kunyudzwa kwechokwadi kwechiratidzo chekare kuderedza kusavimbika. EAGLE-2 inowedzera yakasimba dhizaini muti, uye EAGLE-3 inodonhedza chimiro-chekufanotaura chinomanikidza kuti chiwedzere zvirinani. Sezvineiwo, verification inovimbisa kuburitsa kwakafanana neiyo yakanangwa modhi ingadai yakagadzira yega.
Kucheka latency muvabatsiri vekutaura kuitira kuti mhinduro dzifambe 2-3x nekukurumidza pasina kuchinja mhinduro dzemuenzaniso
Kudzikisa mitengo yeGPU kune yakakwira-vhoriyamu API vanopa nekugadzira akawanda tokeni pakupfuura kwemberi
Kumhanyisa refu ketani-ye-yepfungwa yekufunga modhi apo zviuru zvematokeni zvinogadzirwa pamubvunzo wega wega
Kumhanyisa kodhi-yekupedzisa maturusi uko zvinofungidzirwa, zvinodzokororwa tokeni kutevedzana kunopa yakakwira dhizaini-yekugamuchirwa mitengo.
Yekufungidzira decoding iri kuita default zvivakwa mukushandira stacks sevLLM uye TensorRT-LLM. Tarisira kubatanidzwa kwakasimba ne batching uye KV-cache kugovera, yekuzvinyora modhi iyo isingade yakaparadzana dhizaini, uye Hardware co-dhizaini inotora kufananidzwa kwechokwadi. EAGLE-maitiro ekunyora dhizaini ari kuwedzeredzwa kune akawanda uye ekufunga modhi, uko macheni akareba ekufunga anoita mutengo we-tokeni unonyanya kurwadza, uye kune-chishandiso inference uko latency inonyanya kukosha.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Kudhirodha kwekufungidzira kunomhanyisa modhi yemutauro wakakura nekuita kuti modhi diki diki ifungidzire ma tokeni akati wandei kumberi, ayo modhi hombe yobva yaona mune imwe pass. EAGLE imhando-ye-the-art vhezheni inodhirowa pachikamu chechikamu kwete chechiratidzo, kuendesa 2-4x kumhanya nekurasikirwa zero mumhando yekubuda.
Mudhizaini mudiki anofembera akati wandei kumberi, uye hombe yaivavarirwa modhi inovasimbisa pamwe chete, ichibvuma chivakashure chakareba chakarurama.
EAGLE inofanotaura yakavanzika modhi yemhando uye inoshandisa zvakare musoro wayo weLM, iyo inoburitsa yakanyanya kudhirowa kupfuura nzira dzechiratidzo chete.
Exact speculative sampling preserves the target distribution; this does not mean every random run must emit the same token sequence under different random draws.
Kuisa pane chaiyo yapfuura chiratidzo kunoderedza kusanzwisisika kwekufanotaura chinotevera chakavanzika chimiro, kuvandudza dhizaini kurongeka.
EAGLE-2 yakaunza mamiriro-inoziva dhizaini yemuti, ichiwedzera matavi anovimbisa kukwidza chiyero chekugamuchira.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Kududzira Kwokufungidzira
Tekinoroji