QLoRA na 4-Bit Fine-Tuning
QLoRA bụ usoro na-enye gị ohere ịhazigharị nnukwu ụdị asụsụ na otu GPU ndị na-azụ ahịa site na ịchekwa ụdị oyi kpọnwụrụ na naanị 4 bits kwa arọ.
Nchịkọta
It made customizing 65B-parameter models possible on hardware that previously could only handle models a fraction of that size.
Ime miri emi
Dị ka ọ na-adịkarị, ịmegharị nnukwu ihe nlereanya pụtara ịbuba ibu ọ bụla na 16-bit nkenke yana imelite ha niile, nke chọrọ nnukwu ebe nchekwa. QLoRA na-ejikọta echiche abụọ. Nke mbụ, ọ na-eme ka ihe atụ a zụrụ azụ wee tụọ ya ruo 4 bits, na-ebelata ebe nchekwa ihe dịka okpukpu anọ. Nke abụọ, ọ na-eji LoRA: kama imelite nnukwu matrices dị arọ, ọ na-agbanye obere ihe nkwụnye ọkụ dị ala dị ala n'akụkụ ha, yabụ naanị nde nkeji ole na ole ka emelitere. Ntọala 4-bit ka edobere ebe gradients na-aga naanị site na obere nkwụnye. Ewebatara na 2023 site na Dettmers na ndị ọrụ ibe ya, QLoRA gosipụtara na imezigharị ihe ngosi 65B nke ọma na otu 48GB GPU nwere ike dakọtara n'ịdị mma nke nlegharị anya 16-bit zuru oke.
Nghọta nka nka
QLoRA webatara aghụghọ atọ. NF4 (4-bit NormalFloat) bụ ụdị data emebere maka nkesa mgbịrịgba nke arọ akwara, na-enye izi ezi ka mma karịa int4. Ugboro abụọ quantization compresses quantization constants n'onwe ha, na-azọpụta amara ebe nchekwa. Ndị na-emepe emepe na-eji GPU-CPU jikọtara ọnụ na-amịkọrọ spikes n'oge ogologo usoro, na-egbochi mkpọka na-enweghị ebe nchekwa. N'oge ngafe na-aga n'ihu na azụ azụ, a na-ebelata ihe ọ̀tụ̀tụ̀ 4-bit ka ọ bụrụ 16-bit kpọmkwem n'oge maka matriks mụbaa, wee tụfuo ya.
Mmetụta atụmatụ
Ọsọ na ọnụ ọgụgụ
Usoro ọrụ asụsụ nwere ike ịga ngwa ngwa n'achụghị nkwụsi ike.
Nweta na iru
Ọ na-agbasawanye ohere n'ofe asụsụ na ụdị nzikọrịta ozi.
Mkpebi doro anya
Ndị otu nwere ike itinyekwu oge na ikpe ebe akpaaka na-ejikwa nkwughachi.
Ọdịnihu nke QLoRA na 4-Bit Fine-Tuning
Nlegharị anya nke ọma 4-bit abụrụla omume ọkọlọtọ, nyocha ugbu a na-aga n'ihu na nkenke dị ala, gụnyere ihe nnọchi anya 2-bit na 1-bit (ternary). Atụmatụ nha ọhụrụ dị ka AWQ, GPTQ, na HQQ na-emeziwanye nke ọma, ebe usoro dị ka QA-LoRA na-achọ idobe ihe nlereanya ahụ ọbụlagodi mgbe ha jikọtara ihe nkwụnye ọkụ. Ka ụdị ibu mepere emepe na-eto, na-atụ anya ngwá ọrụ nke na-eme ka ndị na-enwe mmasị na-emegharị 70B-gbakwunyere ụdị n'otu GPU egwuregwu ka ọ bụrụ usoro ihe omume, na-emezi nhazi.
Mmejuputa n'ezie n'ụwa
Mmalite mma na-akụ ụdị 70B Llama n'otu 48GB GPU iji wuo onye na-akwado ndị ahịa n'olu akara nke ya na-enweghị ịgbazite ụyọkọ nkesa.
Onye nyocha nwere otu onye ahịa RTX 4090 na-emegharị ihe ngosi mepere emepe ka ọ bụrụ dataset na-aza ajụjụ ahụike niche n'abali.
Onye nrụpụta na-emepụta ọtụtụ ihe nkwụnye LoRA dị obere, nke nwere ike ịgbanwe maka ọrụ dị iche iche, ha niile na-ekerịta otu ụdị ntọala 4-bit nke etinyere na ebe nchekwa.
Onye na-enwe mmasị na-atọ ụtọ na-akụ ihe nlereanya na ndekọ nkata nke onwe ya iji ṅomie otu ụdị ederede site na iji ngwaike Colab n'efu.
Ihe ize ndụ & okporo ụzọ nche
Eziokwu ndị e chepụtara echepụta nwere ike jiri nwayọ tinye akụkọ, nkwado nkwado, ma ọ bụ nsonaazụ nyocha.
Mmetụta ngwa ngwa nwere ike ịmepụta nsonaazụ na-ekwekọghị ekwekọ n'ofe arịrịọ ndị yiri ya.
Enwere ike ikpughe data ederede nwere mmetụta ma ọ bụrụ na njikwa ohere adịghị ike.
Map mmejuputa
Kọwaa usoro mmepụta, ụda, na ụkpụrụ ịdịmma tupu ibugharị.
Weghachite nzaghachi site na isi mmalite ntụkwasị obi mgbe ọ bụla izi ezi dị mkpa.
Debe ebe nleba anya mmadụ maka mpụta dị elu.
Sochie ụkpụrụ ọdịda ma na-azụghachi mkpali ma ọ bụ usoro ọrụ mgbe niile.
Nọgide na-eme nchọpụta
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.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ntuziaka na-esote
Nlele Nlele Ọjụjụ Dị Mma-Tuning
Ajụjụ a na-ajụkarị
What is QLoRA and 4-Bit Fine-Tuning?
QLoRA bụ usoro na-enye gị ohere ịhazigharị nnukwu ụdị asụsụ na otu GPU ndị na-azụ ahịa site na ịchekwa ụdị oyi kpọnwụrụ na naanị 4 bits kwa arọ. O mere ịhazi ụdị 65B-parameter na ngwaike nke na mbụ nwere ike ijikwa obere akụkụ nke nha ahụ.
Kedu ihe bụ isi nchekwa nchekwa n'azụ akụkụ '4-bit' nke QLoRA?
QLoRA na-echekwa ihe ọ̀tụ̀tụ̀ azụ̀ azụ̀ jụrụ oyi na 4 bit kwa uru, na-ebelata ebe nchekwa ihe dị ka okpukpu anọ ma e jiri ya tụnyere 16-bit.
N'oge nlegharị anya nke QLoRA, kedu paramita ndị a na-emelite site na mgbada gradient?
Ụdị isi bụ oyi kpọnwụrụ; naanị matrices nkwụnye ọkwa dị ala etinyere na-enweta mmelite gradient, nke bụ obere akụkụ nke paramita.
Kedu ihe bụ NF4 n'ihe gbasara QLoRA?
NF4 (NormalFloat 4-bit) bụ ụdị data emebere gburugburu nkesa mgbịrịgba nke ọnụọgụ netwọkụ akwara maka izi ezi karịa int4.
Gịnị bụ nsogbu 'paged optimizers' na adreesị QLoRA?
Ndị na-emepe emepe na-eji GPU-CPU jikọtara ọnụ iji nweta mkpọka ebe nchekwa, na-egbochi mkpọka na-enweghị ncheta n'oge ọzụzụ na ntinye ogologo.
Kedu mgbe a na-atụgharị ihe nha 4-bit laghachi na nkenke dị elu n'oge ọzụzụ?
A na-edobe ịdị arọ 4-bit ka ọ bụrụ nkenke 16-bit na ijiji maka matriks ọ bụla na-amụba, mgbe ahụ, a na-atụfu akwụkwọ nkenke dị elu iji chekwaa ebe nchekwa.