Kediska Isbarbar-dhigga Model iyo Tubooyinka

Jawaabaha iyo sharraxaadaha waxay soo baxayaan markaad qaadanayso kediska.

1. Waa maxay mushkiladda aasaasiga ah ee isbarbar-dhigga qaabka iyo dhuumaha ay xalliyaan?
   • Qaabka oo aad u weyn oo aan ku habboonayn hal xusuusta GPU
   • Xogta aayar ka soo rarista saxanka
   • Gradients qarxinaya inta lagu jiro tababarka
   • Tusaalayaal tababar oo aad u yar ayaa la heli karaa
2. Sidee buu u shaqeeyaa kala qaybsanaanta isbarbardhigga tensor (lakabka gudaha)?
   • Waxay ku meelaysaa lakabyo dhan GPU-yo kala duwan
   • Waxay u kala qaybisaa xisaabta gudaha lakabka, sida matrix ku dhufto, GPU-yada oo dhan
   • Waxay ku koobabaysaa lakab kasta GPU kasta
   • Waxay ku socotaa lakab kasta laba jeer dib-u-celinta
3. Waa maxay 'xumbo' ee isbarbar yaaca dhuumaha naive?
   • Qulqulka xusuusta oo kora waqti ka dib
   • Nuqul nuqul ka mid ah gobolka optimizer
   • Waqti shaqo la'aan ah marka GPU-yada hoose ay sugaan heerarka kor u kaca si ay u soo saaraan hawlqabad
   • Kor u kaca qiyamka jaan-goynta
4. Sidee ayay isbarbar yaaca dhuumaha u yareeyaan xumbo?
   • Adigoo isticmaalaya heerka waxbarashada ka yar
   • Adoo ka saaraya lakabyada qaar ka mid ah moodeelka
   • Adiga oo ku kaydinaya firfircoonida CPU
   • Adiga oo qayb kasta u kala qaybinaya dufcooyin yar-yar oo u gudba marxaladaha si isdaba joog ah
5. Waa maxay sababta isbarbardhigga tensor-ku caadi ahaan loogu hayaa hal nood?
   • Sababtoo ah soo noqnoqoshada dhammaan dhimista natiijooyinka qayb ahaan waxay ka faa'iidaysanaysaa isku xirka qanjirada degdega ah sida NVLink
   • Sababtoo ah uma baahna wax xiriir ah
   • Sababtoo ah nodes ma wadaagi karaan xogta gabi ahaanba
   • Sababtoo ah waxay taageertaa hal GPU oo kaliya
6. Qaabkee ayaa sida ugu wanaagsan loogu yaqaanaa isbarbardhigga tensor-ka ee tababarka beddelka?
   • scikit-barro
   • NVIDIA Megatron-LM
   • FurCV
   • Pandas
7. What data must move between stages in pipeline parallelism during a forward pass?
   • The entire training dataset after each layer
   • Only the optimizer’s learning rate
   • Activation tensors produced by one stage and consumed by the next
   • No data, because stages are independent models
8. Why does tensor parallelism often require fast communication within its group?
   • It assigns each GPU a separate dataset and never synchronizes
   • It communicates only once when the model is loaded
   • It uses no cross-device operations
   • Devices exchange partial results or activations within layers during computation