Ikibazo gihuza ibibazo · Hagati Urwego

Itsinda-Ikibazo Cyitonderwa Ikibazo

Ongera usuzume uburyo itsinda-ryibanze ryitondewe risangira urufunguzo / agaciro imitwe, bigabanya ibisekuruza-igihe cache igiciro, kandi ushobora kuboneka kubigenzura bihari.

Inzira zijyanye no kuyoboraItsinda ryibibazo byitondewe
Ikibazo 1 Bya 8

Mumatsinda-Ikibazo Cyitonderwa, niki gisangiwe mumatsinda yibibazo?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Mumatsinda-Ikibazo Cyitonderwa, niki gisangiwe mumatsinda yibibazo?
    • Urufunguzo n'agaciro imitwe
    • Ibisohoka gusa
    • Iyinjiza
    • Igipimo cyo kwiga
  2. GQA isobanurwa neza nkimpamvu yo hagati hagati yizo ntambwe ebyiri?
    • Kwemeza no kwitabwaho
    • Byuzuye imitwe myinshi hamwe nibibazo byinshi
    • FP32 na FP8 neza
    • Encoder-yonyine na decoder-moderi gusa
  3. Ni ikihe gice cyo gufata umwanzuro GQA ikora cyane cyane?
    • Tokenisation ya progaramu
    • Kubara amahugurwa
    • KV cache yibuka hamwe nubunini bwayo mugihe cyibisekuru
    • Softmax yanyuma hejuru yamagambo
  4. Niba icyitegererezo gifite imitwe 32 yibibazo hamwe nitsinda 8 KV, ubwihisho bwa KV bugabanuka nikihe kintu?
    • Hafi ya 32x
    • Hafi ya 2x
    • Nta kugabanuka
    • Hafi ya 4x
  5. What quality/speed tradeoff did the GQA paper report in its experiments?
    • Quality close to MHA with inference speed closer to MQA
    • Lower quality and slower inference than both
    • Exactly the same latency and quality for every model
    • A result only on image classification
  6. Nigute moderi ihari-imitwe myinshi ihindurwa muri GQA?
    • Mugusiba ibyitonderwa byose
    • Nintambwe ngufi 'kuzamura' kuva kuri bariyeri ihari
    • Muguhindura kuri tokenizer itandukanye
    • Ntibishobora guhinduka na gato
  7. Which task families did the GQA paper evaluate?
    • Image segmentation and object detection
    • Speech synthesis and speaker verification
    • Language modeling and machine translation
    • Molecular generation and toxicity prediction
  8. How much of the original pretraining compute did the GQA paper report using for its uptraining recipe?
    • Abagera kuri 50%
    • About 100%
    • No additional training compute
    • Abagera kuri 5%