Buyela Ezindabeni
UkuqambaAI Understanding ukwaziswa

Iphepha liphakamisa i-CellFill ngezibuyekezo ezingahlehliswa kumamodeli olimi alinganiselwe ngaphandle kokushintsha izingcezu ezigciniwe

Iphepha le-arXiv lethula i-CellFill, indlela enezela ulwazi kumamodeli olimi asetshenzisiwe angu-4-bit ngefayela elihlukile lokubuyekeza kuyilapho ilondoloza izisindo ezibaliwe zangempela. Ababhali babika amanani aphezulu omjovo weqiniso nezibuyekezo ezihlehliswayo, kodwa umsebenzi usalokhu ungowokuphrinta okungakabuyekezwe kuqala odinga…

6 min readRead the primary source
Primary-source image accompanying Paper proposes CellFill for reversible updates to quantized language models without changing stored bits
Idokhumenti yomthombo oyinhlokoUmthombo urekhodiwe
Umshicileli
arxiv.org
Isixhumanisi somthombo
arxiv.orghttps://arxiv.org/abs/2608.20873
Uhlobo lomthombo
Idokhumenti eyisisekelo — isimemezelo esisemthethweni, iphepha, ukugcwalisa, noma ikhasi lomuntu wokuqala esilifunda ngokuqondile.
UmongoQonda lokhu ngemizuzwana engama-60

Qala lapha

Imigomo ebalulekile

Inkumbulo (Inkumbulo yomenzeli)
Ingqikithi egciniwe umenzeli we-AI usebenzisa ezinyathelweni zonke noma izikhathi ukuze athuthukise ukuqhubeka.
Quantization
Iguqula izisindo zemodeli zibe amafomethi anemba aphansi njenge-8-bit noma 4-bit.
Ukulinganisa
Ukuthi amaphuzu okuzethemba kwemodeli afanelana kahle kangakanani namathuba okulutha.
ZihloleImibuzo Ecacisiwe yamamodeli e-AI

Kwenzekeni

Abacwaningi bachaza i-CellFill, indlela yokwengeza ulwazi kumamodeli olimi asetshenzisiwe ngaphandle kokushintsha amakhodi ezinombolo ezigciniwe noma izikali ezabiwe kufayela eliyimodeli yoqobo. Isibuyekezo sigcinwa efayeleni elihlukile ababhali abalibiza ngokuthi gcwalisa futhi lingahoxiswa ngokukhipha.

Iphepha liqala endaweni ye-4-bit . Isisindo semodeli ngayinye simelelwa yinombolo ephelele kanye nesilinganiso esabiwe, kuyilapho inqubo yokulinganisa ilahla isikhawu phakathi kwamanani agciniwe. Ababhali bachaza “ukufunda ngaphakathi kweseli” njengokubhala ulwazi olusha kuleso sikhathi esingasetshenzisiwe. Umkhawulo wabo oshiwo ukuthi ukulinganisa kabusha izisindo ezinikeziwe kufanele kukhiqizwe kabusha amakhodi asekuqaleni nezikali ngokunembile. Ngokwezinto ezibonakalayo, ifayela lemodeli yesisekelo lingahlala lingashintshiwe ezingeni elincane kuyilapho isibuyekezo esihlukile sishintsha ukuziphatha kwemodeli lapho sisetshenziswa.

I-CellFill iwukusetshenziswa kwephepha kwalo mbono. Isebenzisa isikhundla sezinga eliphansi ngaphakathi kweseli ngalinye lokulinganisa futhi iqeqeshelwa ukukhishwa komthengisi okungu-4-bit. Umthombo uthi ukuhlolwa kuhlanganisa i-NF4, i--aware-training kanye negridi yesitayela se-GPTQ, kanye namamodeli we-Qwen3 kanye ne-Gemma asukela kupharamitha eyizigidi eziyizinkulungwane eziyi-1.7 kuya kwezingu-31. Ababhali bathi le ndlela ihlola isiqinisekiso sayo esizindeni esiphelele saso sonke isisindo esiboshiwe, kunokuthembela kuphela ezilinganisweni zezinombolo ezihlanganisiwe.

Ababhali babika ukuthi i-CellFill ibhale u-83% kuya ku-97% wekhophasi yamaqiniso angempela amamodeli ahloliwe ayengawazi ngokuqinisekile, kanti kuneziro ezibikiwe zokwephulwa kwezingqinamba kuzo zonke izisindo eziyi-2.4 × 10^10 ezivinjiwe. Baphinde babike izivivinyo okuhloswe ngazo ukukhombisa ukuthi ulwazi olwengeziwe lungasetshenziswa kunokuba kukhunjulwe nje: izidakamizwa ezijovwe zingahlanganiswa nezithako ezijovwe kuze kufike kusilingi obikiwe wemodeli we-hop emibili, kanye nomtapo wezincwadi wesofthiwe okhishwe ngemva kokunqanyulwa kwemodeli kusetshenziswe ikhodi ngethuba lezikhathi ezingu-2.6. Lezi izimangalo ezivela ekuhlolweni kwangaphambilini, hhayi imiphumela esungulwe ngokuzimela.

Ukuze uthole ukuqhathanisa ukubuyisela emibuzweni yomsila omude we-PopQA, umthombo uthi ukugcwaliswa kuphendule imibuzo engu-82% kuya ku-90% imodeli ekhishwe engakwazi ukuyiphendula, isebenzisa amathokheni angu-75 ngombuzo ngamunye. Amasistimu okubuyisa acashuniwe asebenzise amathokheni angama-90 kuye kwayi-1,008 ngombuzo ngamunye. Leli phepha liphinde lithi ababhali abaningana bangabelana ngokukhishwa kwemodeli eyodwa nokuthi izibuyekezo ezilandelanayo zisebenzisa ingxenye engashintshi yegumbi elisele. Umthombo ukhomba kumafayela emiphumela afakwe kungobo yomlando ngemuva kwamathebula awo, kodwa umbhalo onikeziwe awuchazi okuqukethwe kwawo ngemininingwane eyanele ukuze kuhlolwe umthetho olandelwayo wokuhlola ogcwele.

Imininingwane yomthombo: arxiv.org ↗

Kungani kubalulekile

Uma iqinisekiswa ngokuzimele, indlela yokwenza ingenza ezinye izibuyekezo zemodeli kube lula ukuzicwaninga, ukusatshalaliswa, ukuqinisekiswa nokuhoxiswa ngoba ukukhishwa kwasekuqaleni kuhlala kufana kancane kuyilapho ukuziphatha okusha kwenziwa ngokuhlukana. Umthombo ubika imiphumela ethembisayo, kodwa awuqinisekisi ukulungela ukukhiqiza noma ukwethembeka okubanzi.

Ukubaluleka okusebenzayo ukulawula inguqulo. Imodeli yolimi esetshenzisiwe ingase ixhunywe emibikweni yokuma, izitifiketi, inqwaba yamadivayisi noma amanye amasistimu athatha ifayela elithile. Isibuyekezo esijwayelekile singadala i-artifact yemodeli entsha futhi senze kube nzima ukuqhathanisa nokukhishwa okuqinisekisiwe. Ukwehlukaniswa okuhlongozwayo kwe-CellFill phakathi kokukhishwa kwesisekelo okungashintshiwe kanye nokugcwaliswa okukhishwayo kungavumela inhlangano ukuthi isabalalise ulwazi olusha ngaphandle kokufaka esikhundleni sefayela langempela elilinganisiwe. Lokho kuhlukana kungenza ukuguqulwa kube sobala kakhulu. Ngokusho kwephepha, isibuyekezo siyifayela elihlukile elingahoxiswa ngokukhipha. Lokhu kudala umehluko ongaba usizo phakathi kwemodeli yesisekelo eqinile kanye nesendlalelo sokubuyekeza esingangezwa, sisuswe noma sabelwe abasebenzisi abahlukene. Idizayini enjalo ingase isize o-opharetha ukulandelela ukuthi yiluphi ulwazi olungeziwe nokuthi nini, nakuba umthombo ungakwethuli ukucwaninga okuphelele, ukulawula ukufinyelela noma uhlelo lokuphatha.

Ukuqhathanisa ukusebenza kahle okubikiwe kukhomba kwenye inzuzo engaba khona. Ukugcwaliswa kuphendule imibuzo eminingi ye-PopQA ebiphuthelwe ngaphambilini ngesabelomali sethokheni esifushane kakhulu kunezinhlelo zokubuyisa ezisetshenziswe ekuqhathaniseni. Uma umphumela usinda ekuphindaphindeni futhi usebenza kumthwalo obanzi womsebenzi, isibuyekezo esihlangene kwesinye isikhathi singabiza noma kube lula ukusetshenzwa kunokubuyisa umongo omude ngokuphindaphindiwe. Umthombo awunikezi ulwazi olwanele ukuze unqume ukuthi izilawuli zokuqhathanisa zenkomba, ikhwalithi yokubuyisa, ukubambezeleka, ukusetshenziswa kwememori noma izindleko zesistimu ziphelele.

Indlela futhi ishintsha indlela izibuyekezo zemodeli ezingase zihlolwe ngayo. Ngenxa yokuthi ababhali badinga ukulondolozwa okunembile kwamakhodi nezikali ezibaliwe, banikeza ukhonkolo, isimo esingahloleka ngomshini se-artifact eyisisekelo. Lokho kungenza izimangalo ezithile zokubuyekeza kube lula ukuqinisekiswa kunezimangalo ezijwayelekile zokulungisa kahle. Kodwa-ke, i-bit identity ayifani nokuphepha kokuziphatha noma ikhwalithi: imodeli ingaziphatha ngendlela ehlukile ngokugcwaliswa ngisho noma izingcezu zayo eziyisisekelo ezigciniwe zingashintshiwe. Umthombo awubonisi ukuthi le ndlela iyavikela yini ekuboneni izinto ezingekho, ukuzihlanganisa okungafunwa, ukuvuza kwedatha noma ulwazi olusanda kujovwa oluyingozi.

Interactive Mechanism

I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela

Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.

Document Size:128K tokens
Needle Placement Depth (Location in document):50% into text
Attention Context Buffer Map:
Target Fact (50%)
Equivalent Pages~320Standard book pages
Retrieval Accuracy99.9%Needle recall score
RAM / KV Cache5.1 GBMemory overhead
Prompt CachingActive~80% discount on reuse
Core takeaway: Million-token context windows allow querying whole codebases or legal archives in one prompt. However, KV cache memory scales with context length, making prompt caching crucial for real-time production.
I-Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

Ongakubuka ngokulandelayo

Imibuzo eyinhloko ukuthi ingabe abacwaningi bangaphandle bangakwazi yini ukukhiqiza kabusha imiphumela ebikiwe, noma ngabe ukugcwalisa ukulondoloza amandla akhona, ukuthi izibuyekezo eziningi zisebenzisana kanjani, nokuthi singakanani isitoreji nesikhathi sokusebenza esingeziwe indlela eyengeza ngayo. I-preprint iphinda ishiye imibuzo evulekile mayelana nokuphepha, ukusetshenziswa kabi kanye nokusebenza ngale kwezilungiselelo ezihloliwe.

Ukuphindaphinda okuzimele kuwuhlolo lokuqala oluyisihluthulelo. Leli phepha liyi-arXiv eliphrintiwe ngaphambili, elethwe ngomhla ka-21 Agasti 2026 futhi labuyekezwa ngomhla ka-24 Agasti 2026, esikhundleni sokushicilelwa kontanga kwirekhodi elinikeziwe. Abacwaningi bazodinga ukuthi baqinisekise isiqinisekiso sesizinda esiphelele, umphumela wokwephulwa kweqanda kanye namazinga omjovo wolwazi abikiwe kuzo zonke izikimu zokulinganisa kanye nemindeni eyimodeli. Amafayela emiphumela afakwe kungobo yomlando angasiza, kodwa umbhalo womthombo wodwa awuqinisekisi ukufinyeleleka kwawo, ukuphelela noma ukuqinisekiswa okuzimele.

Ukuhlola okuzayo kufanele kuhlole ukuthi amandla okugcwaliswa ayalimaza yini imodeli eyisisekelo ebisivele inawo. I-abstract enikeziwe igcizelela amaqiniso amamodeli abengawazi, ukuhlolwa kokuqamba kanye nelabhulali yesofthiwe yangemuva kokunqanyulwa, kodwa ayibiki ukuhlolwa kokuhlehla okubanzi, ukuziphatha kokunqaba, ukulinganisa, ukuba yiqiniso ngaphandle kwekhorasi ejovwe noma ukusebenza emisebenzini engahlobene. Futhi ayichazi ukuthi ababhali banqume kanjani ukuthi amamodeli ashoda ngeqiniso ngalinye ngaphambi kokuqeqeshwa.

Izimangalo zephepha mayelana nezibuyekezo ezilandelanayo nababhali abaningi zidinga ukuhlolwa ngokucophelela. Indlela esebenzela ukugcwaliswa okukodwa ingase iziphathe ngendlela ehlukile njengoba kusetshenziswa isikhala seseli esitholakalayo, njengoba izibuyekezo zigqagqana noma njengoba ababhali bethula amaqiniso angqubuzanayo. Izilinganiso ezibalulekile ezisebenzayo zifaka usayizi wokugcwalisa ngakunye, ukuqeqeshwa nesikhathi sokufaka isicelo, ukubambezeleka kwemibono, ukusetshenziswa kwenkumbulo, inani lezibuyekezo ezisimeme kanye nezimo lapho ukususa kubuyisela ngokugcwele ukuziphatha kwangaphambilini.

Ezokuphepha nokubusa nakho akukaxazululwa. Ifayela lesibuyekezo elikhiphekayo lingaba usizo ekusetshenzisweni okulawulwayo, kodwa lingase libe impokophelo ezimele yokuphazamisa uma umhlaseli engakwazi ukulishintsha noma alishintshe. Umthombo awuchazi ukuqinisekiswa, ukugunyazwa, ukuvela, ukuvikelwa kobumfihlo noma ukuzivikela ekujovweni kolwazi olunonya. Buka ukuhoxiswa kokulandelela kokuhlolwa komsebenzi kumasistimu wokupha wangempela, ukusebenzisana nokushuna kokuphepha nokusebenza kumamodeli, izizinda nezinhlobo zokubuyekeza ngale kokuhlolwa okubikwe lapha.

Imihlahlandlela ehlobene nemibuzo

Amamodeli e-AI AchaziweChatGPT ne-LLMsUkuqeqeshwa kwe-AIUkuziphatha kwe-AIHlola okwaziyo — zama imibuzo ye-AI yamahhalaBheka igama le-AI kuhlu lwethu lwamagamaLandela i-tracker yokukhishwa kwemodeli ye-AI
Uthole lokhu kuwusizo?