Jagoran Harshe AI

QLoRA da 4-Bit Fine-Tuning

QLoRA wata dabara ce da ke ba ku damar daidaita ɗimbin ƙirar harshe akan GPU mabukata guda ɗaya ta hanyar adana daskararrun ƙirar a cikin rago 4 kawai a kowane nauyi.

2 min karatuAn sabunta ta ƙarshe

Dubawa

It made customizing 65B-parameter models possible on hardware that previously could only handle models a fraction of that size.

Zurfafa nutsewa

A al'ada, daidaitawa babban samfuri yana nufin ɗaukar kowane nauyi a daidaitaccen 16-bit da sabunta su duka, wanda ke buƙatar babban ƙwaƙwalwar ajiya. QLoRA ya haɗu da ra'ayoyi biyu. Da farko, yana daskare samfurin da aka riga aka horar kuma yana ƙididdige shi zuwa 4 ragowa, yana kashe ƙwaƙwalwar ajiya kusan ninki huɗu. Na biyu, yana amfani da LoRA: maimakon sabunta manyan matrices masu nauyi, yana allura ƴan ƙaramin adaftar adaftar ma'auni tare da su, don haka kawai ƴan sigogi miliyan ne ake sabunta su. Tushen 4-bit yana tsayawa yayin da gradients ke gudana ta cikin ƙananan adaftan kawai. An gabatar da shi a cikin 2023 ta Dettmers da abokan aiki, QLoRA ya nuna cewa ingantaccen tsarin 65B akan GPU 48GB guda ɗaya zai iya dacewa da ingancin cikakken daidaitawar 16-bit mai kyau.

Fahimtar Fasaha

QLoRA ya gabatar da dabaru uku. NF4 (4-bit NormalFloat) nau'in bayanai ne wanda aka inganta don rarraba kararrawa na ma'aunin jijiyoyi, yana ba da daidaito mafi inganci fiye da bayyanannen int4. Ƙididdigar ƙididdigewa sau biyu tana damfara ma'aunin ƙididdigewa da kansu, yana adana ƙarin ƙwaƙwalwar ajiya. Masu inganta shafi na yin amfani da haɗewar ƙwaƙwalwar ajiya ta GPU-CPU don ɗaukar spikes yayin dogon jeri, da hana faɗuwar abin tunawa. Yayin wucewar gaba da baya, ma'aunin 4-bit ana ƙididdige su zuwa 16-bit daidai-in-lokaci don haɓaka matrix, sannan a jefar da su.

Dabarun Tasiri

Gudu da sikelin

Gudun aikin harshe na iya tafiya da sauri ba tare da sadaukar da daidaito ba.

Shiga ku isa

Yana faɗaɗa damar shiga cikin harsuna da salon sadarwa.

Shawarwari masu haske

Ƙungiyoyi za su iya ciyar da ƙarin lokaci akan hukunci yayin da aiki da kai ke sarrafa maimaitawa.

Makomar QLoRA da 4-Bit Fine-Tuning

4-bit fine-tuning ya zama daidaitaccen aiki, kuma bincike yanzu yana turawa zuwa madaidaicin madaidaici, gami da wakilcin 2-bit da 1-bit (ternary). Sabbin tsare-tsare na ƙididdigewa kamar AWQ, GPTQ, da HQQ suna ƙara inganta daidaito, yayin da dabaru irin su QA-LoRA suna nufin kiyaye ƙirar ƙididdigewa koda bayan haɗa adaftar. As open-weight models grow, expect tooling that lets hobbyists fine-tune 70B-plus models on a single gaming GPU to become routine, democratizing customization.

Aiwatar da Gaskiyar Duniya

Farawa mai kyau-sautun samfurin 70B Llama akan GPU guda 48GB don gina mataimaki na goyon bayan abokin ciniki a cikin muryar saƙon sa ba tare da hayar tarin sabar ba.

Mai bincike tare da mabukaci guda ɗaya RTX 4090 yana daidaita ƙirar buɗaɗɗe zuwa saitin bayanan amsa tambayoyin likita na dare.

Mai haɓakawa yana ƙirƙirar ƙanana, masu adaftar LoRA masu musanyawa don ayyuka daban-daban, duk suna raba ƙirar tushe guda 4-bit wanda aka loda cikin ƙwaƙwalwar ajiya.

Masu sha'awar sha'awa suna tsara samfuri akan rajistan ayyukansu na sirri don kwaikwayi wani salon rubutu ta amfani da kayan aikin Colab kyauta.

Hatsari & Tsare-tsare

Abubuwan da aka ruɗe suna iya shigar da rahotanni cikin nutsuwa, kwararar tallafi, ko abubuwan bincike.

Hankali na gaggawa na iya ƙirƙirar sakamako mara daidaituwa a cikin buƙatun iri ɗaya.

Za a iya fallasa bayanan rubutu mai ma'ana idan ikon samun dama yana da rauni.

Taswirar Hanya

1

Ƙayyade tsarin fitarwa, sautin, da ma'auni masu inganci kafin fitowa.

2

Amsa a ƙasa tare da amintattun tushe a duk lokacin da daidaito ya shafi mahimmanci.

3

Ajiye wurin binciken ɗan adam don abubuwan da ake samu masu girma.

4

Bibiyar tsarin gazawar kuma sake horar da tsokaci ko tafiyar aiki akai-akai.

Ci gaba da Bincike

Free newsletter

Get the daily AI briefing

Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.

One email each weekday. Unsubscribe in one click. We never sell or share your address.

Test yourself

Take the QLoRA and 4-Bit Fine-Tuning quiz

Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.

Fara tambayoyi

Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation

Jagora na gaba

Ƙimar Samfura Mai Kyau-Tuning

Tambayoyin da ake yawan yi

What is QLoRA and 4-Bit Fine-Tuning?

QLoRA wata dabara ce da ke ba ku damar daidaita ɗimbin ƙirar harshe akan GPU mabukata guda ɗaya ta hanyar adana daskararrun ƙirar a cikin rago 4 kawai a kowane nauyi. Ya sanya keɓance samfuran siga-65B mai yiwuwa akan kayan masarufi waɗanda a baya zasu iya ɗaukar nau'ikan juzu'in girman wannan.

Menene ainihin ra'ayin ceton ƙwaƙwalwar ajiya a bayan ɓangaren '4-bit' na QLoRA?

QLoRA tana adana ma'aunin daskararrun da aka riga aka horar da su a 4 ragi a kowace ƙima, yana yanke ƙwaƙwalwar ajiya kusan ninki huɗu idan aka kwatanta da 16-bit.

Yayin daidaitawa mai kyau na QLoRA, wadanne sigogi ne ake sabunta su ta hanyar zuriyar gradient?

Tsarin tushe yana daskarewa; kawai alluran matrix adaftar ma'auni kaɗan ne ke karɓar sabuntawar gradient, wanda ƙaramin juzu'i ne na sigogi.

Menene NF4 a cikin mahallin QLoRA?

NF4 (NormalFloat 4-bit) nau'in bayanai ne da aka tsara a kusa da rarraba kararrawa na ma'aunin cibiyar sadarwa don ingantacciyar daidaito fiye da bayyanannen int4.

Wace matsala 'masu inganta shafi' a adireshin QLoRA?

Masu inganta shafi na yin amfani da haɗe-haɗen ƙwaƙwalwar ajiya na GPU-CPU don ɗaukar ƙaƙƙarfan ƙaƙƙarfan ƙwaƙwalwar ajiya, da hana faɗuwar faɗuwar ƙwaƙwalwa yayin horo akan dogon bayanai.

Yaushe ma'aunin 4-bit ke juyawa zuwa mafi girman daidaito yayin horo?

An ƙididdige ma'aunin 4-bit zuwa madaidaicin 16-bit akan tashi don kowane matrix ninka, sannan ana jefar mafi girman kwafin don adana ƙwaƙwalwar ajiya.