Overview
Positional Interpolation (PI) inzira yakapfava, ine pesvedzero inotambanudza hwindo reTransformer remukati nekudzvanya ma indices ezvinzvimbo muhuwandu iyo modhi inotoziva. Panzvimbo pokuwedzera kune zvisingaoneki, inopindirana mukati mevakadzidziswa, ichingoda kurongeka kwenguva pfupi.
Positional Interpolation for Long Context inyanzvi yekuvaka inobata mhando yemhando, mutengo wezvivakwa, latency, uye kuvimbika pachiyero.
Deep Dive
Yakaunzwa ne Meta vaongorori (Chen et al.) muna 2023, Positional Interpolation inobata chokwadi chekuti mamodheru ane RoPE anokundikana zvakanyanya kana achiwedzera kune zvinzvimbo kupfuura kudzidziswa. Muono wacho unopokana: pane kukumbira modhi kuti ibate mahombe ezvinzvimbo zvayasati yamboona, PI inokamura maindisi enzvimbo anouya nechikero kuitira kureba kwe, toti, 8K mepu kudzoka muyekutanga 2K renji. Nekuti iyo modhi yakadzidziswa pane iyo siyana, iyo inotenderera inogara mu-kugovera. Mushure mekungoita chiuru-yekugadzirisa matanho, modhi yeLLaMA yakawedzera nzira iyi yakabatwa kusvika kumakumi matatu nemaviri mamiriro. Iro bepa rakaratidza kuwedzeredza kunogona kukwidza kutarisisa zvibodzwa kune hukuru hukuru, nepo kududzira kuchiita kuti irambe yakasungwa uye yakagadzikana, ndosaka kududzira kunoshanda zvakanyanya zvirinani pane kuwedzera.
Technical Insight
PI inodzoreredza nzvimbo m kusvika m/s uko s iri chinhu chekuwedzera (semuenzaniso, hurefu hutsva hwakakamurwa nehurefu hwepakutanga). KuRoPE izvi zvinonyatso dzikisa nhanho yekutenderera pakati pezvinzvimbo zvakapakana, kurongedza mamwe zvinzvimbo muakadzidziswa angular renji. Iyo theoretical yakasungwa mubepa inoratidza kutarisisa kwakaringana zvibodzwa zvinogara zvakadzorwa, nepo naive extrapolation inogona kuburitsa maodha ehukuru hwakakura kupfuura chero chinhu chinoonekwa mukudzidziswa, kudzikamisa softmax.
Mastering Positional Interpolation for Long Context
Positional Interpolation (PI) inzira yakapfava, ine pesvedzero inotambanudza hwindo reTransformer remukati nekudzvanya ma indices ezvinzvimbo muhuwandu iyo modhi inotoziva. Panzvimbo pokuwedzera kune zvisingaoneki, inopindirana mukati mevakadzidziswa, ichingoda kurongeka kwenguva pfupi. Positional Interpolation for Long Context inyanzvi yekuvaka inobata mhando yemhando, mutengo wezvivakwa, latency, uye kuvimbika pachiyero. Kuvaka kunzwisisa kwakadzama, tora Positional Interpolation for Long Context semuenzaniso wekushandisa, kwete chinhu chimwe chete: tsanangura zvinodiwa, kujekesa fungidziro, uye patsanura izvo zvingaitwe nehurongwa hwakavimbika kubva kune zvichiri kuda kutonga kwenyanzvi.
Mukuita, zvikwata zvakasimba zvinoshandisa Positional Interpolation for Long Context 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
Kuwedzera 2K-context LLaMA modhi yekubata 8K-32K tokens ine angangoita chiuru-yekugadzirisa matanho.
Kugadzirisa iyo iripo yekukurukura modhi yekupfupisa-refu-gwara pasina kudzidziswa kubva pakutanga
Kushanda seyambiro yekutanga iyo NTK-inoziva kuyera uye YaRN inovandudza pairi
Kugonesa kodhi-yakareba kodhi kana yemutemo-gwaro kuongororwa pamhando dzakadzidziswa pakutanga nemahwindo mapfupi
Maitiro Ekuita
Positional Interpolation for Long Context mukuita
Kuwedzera 2K-context LLaMA modhi yekubata 8K-32K tokens ine angangoita chiuru-tuning nhanho.
Kuwedzera 2K-context LLaMA modhi yekubata 8K-32K tokens ine angangoita chiuru-tuning matanho Matimu anowanzo kuwana mhedzisiro iri nani kana achinge atsanangura emhando yepamusoro kumberi, chengeta nzira yekukwira kwevanhu yemakesi emupendero, uye kuteedzera zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Positional Interpolation for Long Context mukuita
Kugadzirisa iyo iripo yekukurukura modhi yekupfupisa-refu-gwara pasina kudzidziswa kubva pakutanga.
Kugadzirisa iripo yekukurukura modhi yekupfupisa-refu-zvinyorwa pasina kudzidziswa kubva kukutanga 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.
Positional Interpolation for Long Context mukuita
Kushanda seyambiro yekutanga iyo NTK-inoziva kuyera uye YaRN inovandudza pairi.
Kushanda seyambiro yekutanga iyo NTK-inoziva kuyera uye YaRN inonatsiridza paMatimu anowanzo kuwana zvirinani zvibodzwa kana vachitsanangudza hunhu hwepamberi, chengetedza nzira yekukwira kwevanhu yemakesi emupendero, uye tarisa zvese zvakawanikwa zvechigadzirwa uye mutengo wekukanganisa nekufamba kwenguva.
Positional Interpolation for Long Context mukuita
Kugonesa kodhi-yakareba kodhi kana yemutemo-gwaro kuongororwa pamhando dzakadzidziswa pakutanga nemahwindo mapfupi.
Kugonesa kodhi-yakareba kodhi kana pamutemo-gwaro kuongororwa pamamodhi akadzidziswa mapfupi windows Zvikwata zvinowanzowana mhedzisiro iri nani kana vachitsanangudza zvikumbaridzo zvemhando kumberi, chengetedza nzira yekukwira kwevanhu yemakesi ekumucheto, 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.