Overview
Panzvimbo pekufanotaura chete chiratidzo chinotevera, modhi inodzidziswa kufanotaura akawanda emangwana tokeni kamwechete. Izvi zvinorodza masaini ekudzidza uye kuvhura kukurumidza kufungidzira kuburikidza nekuzvifungira decoding.
Multi-Token Prediction Training chikamu chemutauro-AI stack inoshandiswa kuverenga, kugadzira, kuronga, uye kushandura zvinyorwa uye kutaura pamwero.
Deep Dive
Mienzaniso yemitauro yakajairwa inodzidziswa nekufanotaura kunotevera-chiratidzo: kupihwa mamiriro, fungidzira chiratidzo chinotevera. Multi-token prediction (MTP), yakafarirwa nebepa ra2024 Meta uye yakagamuchirwa muDeepSeek-V3, inowedzera mimwe misoro yakareruka yekubuda kuitira kuti modhi yacho ifanofembera panguva imwe chete chiratidzo chinotevera pamwe neyechipiri, 3, uye 4th tokens kumberi kubva munzvimbo imwechete yakavanzika. Izvi zvinomanikidza network kuronga mberi mune ramangwana uye inosimbisa chiratidzo chekudzidzisa - imwe neimwe nzvimbo ikozvino inopa akawanda kurasikirwa mazwi. Meta yakashuma zvakawandirwa zvakanyanya pakunyora kodhi uye kufunga kunovaka, nemhando dzakakura dzichibatsira zvakanyanya. Zvine hutsinye, iyo misoro yekuwedzera inogona kuraswa mushure mekudzidziswa, saka saizi yemhando pakutumirwa haifanire kukura.
Technical Insight
MTP inonamatira n yakazvimiririra yekufanotaura misoro pamusoro peiyo yakagovaniswa transformer trunk; musoro k unofanotaura chiratidzo pachinzvimbo t+k kubva pane chinomiririra pachinzvimbo t. Kurasikirwa kunopfupikiswa panguva yekudzidziswa. Pakunongedza, misoro yebetsero inogonesa kuzvifungira decoding: modhi inokurudzira tokeni akati wandei mukupasa kumwe chete, wobva wazvisimbisa, uchiwana kusvika kunosvika 3x inokurumidza chizvarwa pasina kushandura kugovera.
Mastering Multi-Token Prediction Training
Panzvimbo pekufanotaura chete chiratidzo chinotevera, modhi inodzidziswa kufanotaura akawanda emangwana tokeni kamwechete. Izvi zvinorodza masaini ekudzidza uye kuvhura kukurumidza kufungidzira kuburikidza nekuzvifungira decoding. Multi-Token Prediction Training chikamu chemutauro-AI stack inoshandiswa kuverenga, kugadzira, kuronga, uye kushandura zvinyorwa uye kutaura pamwero. Kuti uvake kunzwisisa kwakadzama, tora Multi-Token Prediction Training semuenzaniso wekushandisa, kwete chinhu chimwe chete: tsanangura zvinodikanwa, kujekesa fungidziro, uye patsanura zvinogona kuitwa nehurongwa hwakavimbika kubva kune zvichiri kuda kutonga kwenyanzvi.
Mukuita, zvikwata zvakasimba zvinoshandisa Multi-Token Prediction Training dhizaini zvinokurudzira, kudzoreredza, uye kuongorora zvishwe seimwe yakabatanidzwa yekutaurirana system. 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.
Mutauro workflows inogona kufamba nekukurumidza pasina kupira kuenderana. Panguva imwecheteyo, chokwadi cheHallucified chinogona kupinda chinyararire mishumo, kuyerera kwetsigiro, kana kutsvagisa zvinobuda. 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
Mutauro workflows inogona kufamba nekukurumidza pasina kupira kuenderana.
Mutauro workflows inogona kufamba nekukurumidza pasina kupira kuenderana. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Inopamhidzira kupinda mumitauro yese nemataera ekutaurirana.
Inopamhidzira kupinda mumitauro yese nemataera ekutaurirana. Mukutumirwa kwemhando yepamusoro, izvi zvinoshandurirwa kuita mitemo inoyerwa yekushanda, miganhu yevaridzi, uye tsika dzekudzokorora dzinodzokororwa kuitira kuti zvikwata zvikwire kuvimba pane kukwidza kusajeka.
Zvikwata zvinogona kupedza nguva yakawanda pakutonga uku otomatiki ichibata kudzokorora.
Zvikwata zvinogona kupedza nguva yakawanda pakutonga uku otomatiki ichibata kudzokorora. 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
DeepSeek-V3 uchishandisa chinangwa cheMTP panguva yekudzidzira kuwedzera kushanda kwedata uye kugonesa kufungidzira decoding.
Meta's macode-generation modhi anoratidza huchokwadi hunowanikwa paHumanEval neMBPP kubva mukufanotaura tokeni dzakawanda.
Kuzvifungira wega decoding: kunyora 3-4 tokens per forward pass wobva waona kukurumidza, kugovera-kuchengetedza kubuda.
Inokurumidza kuzadzisa otomatiki mumakodha vabatsiri uko akawanda anonzwisisika tokeni anotsanangurwa uye anotariswa mune imwe nhanho.
Maitiro Ekuita
Multi-Token Prediction Training mukuita
DeepSeek-V3 uchishandisa chinangwa cheMTP panguva yekudzidzira kuwedzera hunyanzvi hwedata uye kugonesa kufungidzira decoding.
DeepSeek-V3 ichishandisa chinangwa cheMTP panguva yekudzidzira kukwidziridza hunyanzvi hwedatha uye kugonesa fungidziro yekudhirodha Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura hunhu hwepamberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Multi-Token Prediction Training mukuita
Meta's macode-generation modhi anoratidza huchokwadi hunowanikwa paHumanEval neMBPP kubva mukufembera tokeni dzakawanda.
Meta's macode-generation modhi anoratidza chokwadi chaakawana paHumanEval neMBPP kubva mukufembera akawanda tokens Matimu anowanzowana mibairo iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa nemitengo yekukanganisa nekufamba kwenguva.
Multi-Token Prediction Training mukuita
Kuzvifungira-decoding: kunyora 3-4 tokens per forward pass wobva waona kukurumidza, kugovera-kuchengetedza kuburitsa.
Kuzvifungira wega decoding: kunyora 3-4 tokens per front pass wobva waona kukurumidza, kugovera-kuchengetedza zvinobuda Matimu anowanzo kuwana mibairo iri nani kana achinge atsanangura hunhu hwepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye kukanganisa mutengo nekufamba kwenguva.
Multi-Token Prediction Training mukuita
Inokurumidza kuzadzisa otomatiki mumakodha vabatsiri uko akawanda anonzwisisika tokeni anotsanangurwa uye anotariswa mune imwe nhanho.
Kurumidza kupedzisa otomatiki mumakodha vabatsiri uko akawanda anonzwisisika tokeni anotsanangurwa uye kutariswa mune imwe nhanho Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Njodzi & Guardrails
Chokwadi chehuroyi chinogona kupinda chinyararire mishumo, kuyerera kwetsigiro, kana tsvakiridzo.
Kunzwa nekukasira kunogona kugadzira mhedzisiro isingaenderane pane zvikumbiro zvakafanana.
Sensitive text data inogona kuburitswa kana zvidhiraivho zvisina kusimba.
Implementation Roadmap
Tsanangura chimiro chekubuda, toni, uye mhando zviyero usati waburitsa.
Tsanangura chimiro chekubuda, toni, uye mhando zviyero usati waburitsa. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Mhinduro dzepasi neakavimbika masosi pese pazvine basa.
Mhinduro dzepasi neakavimbika masosi pese pazvine basa. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Chengetedza ongororo yekuongorora yemunhu kune yakakwira-stake zvinobuda.
Chengetedza ongororo yekuongorora yemunhu kune yakakwira-stake zvinobuda. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.
Tevera maitiro ekutadza uye dzidzisazve kukurudzira kana mafambiro ebasa nguva nenguva.
Tevera maitiro ekutadza uye dzidzisazve kukurudzira kana mafambiro ebasa nguva nenguva. Bata nhanho yega yega segedhi rehumbowo: kana maitiro asina kusangana, imbomira kuburitsa, vhara gaka, uye wobva wawedzera kushandiswa.