quiz bu lëkkale ak tegtal · Diggante Niveau

QLoRA ak laaj ak tontu bu 4-bit

Xoolaat base 4-bit bu QLoRA, adaptatër yu rang yu woyof yuñ mëna tàggat, NF4, ñaari kantifikaasioŋ, optimisatëri paged, ak dekantisaasioŋ bu yàgg.

Yooni njiit yu jëm ciQlora Ak 4 Bit
Laaj 1 de 8

Lan mooy xalaatu sakkanal mémoire bi ci ginaaw wàllu '4-bit' ci QLoRA?

Xoolaatal laaj bu nekk8 laaj

Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.

  1. Lan mooy xalaatu sakkanal mémoire bi ci ginaaw wàllu '4-bit' ci QLoRA?
    • Dafay xayma diisaayu modelu base bu liw ba 4 bit ci bu nekk
    • Dafay efaase 75% ci diisaayi model bi
    • Dafay tàggat model bi ci 4 jamono kese
    • GPUs rek lay jëfandikoo, du bari
  2. Ci diiru ajustement QLoRA, ban parametre lañuy yeesal ci wàccinu gradient bi?
    • Diisaay yiñ njëkka am ci model bi yépp
    • matris adaptatëru LoRA yu ndaw yi am rang bu woyof kese
    • Konstantu kantifikaasioŋ bu 4 bit
    • Baatu tokenizer bi
  3. Luy NF4 ci wàllu QLoRA?
    • Reseau neuronal bu am 4 couche
    • Algorithm buy gëna xéewale
    • Xeetu done bu am 4 bit buñ jagleel poid yiñ séddale ci anam wu jaar yoon
    • Format GPU bu am 4 node
  4. Ban jafe-jafe la 'optimiser paged' yi ci QLoRA di saafara?
    • Yebbi fichier model yu yeex ci internet
    • Tokenisation bu jaarul yoon
    • Overfitting ci done yu ndaw
    • Pike yu génn ci mémoire ci diir yu yàgg
  5. Kañ lañuy soppi poid 4-bit yi ñu delloo ci gëna dëggu ci diiru tàggat yaram?
    • Just-in-time ci diiru bariwaayu matrix bi, ba noppi sànni ko
    • Duñu musa tuub; math ñu ngi koy defee ci 4 bit
    • Benn yoon kese ci ndoorte tàggat yaram
    • Ginaaw bi tàggat bi jeexee
  6. Lan la këyitu QLoRA bi njëkk wane ci wàllu kalite?
    • 4-bit fine-tuning dafay raw fine-tuning bu mat sëkk
    • Reglage bu baax ci benn model 65B ci benn GPU mën na méngoo ak kalite reglage bu 16-bit
    • Kantitasioŋ dafay tax model yi amul njariñ ci suufu 8 bit
    • Modèle yu ndaw yi kese lañu mëna defar ci anam yii
  7. What does QLoRA’s double-quantization step compress?
    • Baatu tokenisatër bi
    • The LoRA adapter rank
    • The quantization constants used to represent the 4-bit weights
    • The training examples’ text labels
  8. How can one QLoRA base model support multiple task-specific versions?
    • They delete the optimizer state after every token
    • They increase the base model from 4-bit to 32-bit
    • They move all activations into the tokenizer
    • Reuse the frozen 4-bit base and swap small trainable LoRA adapters for different tasks