quiz bu lëkkale ak tegtal · Dëgër Niveau
Block-Sparse ak laaj ak tontu bu bari
Saytu blok-sparse bàyyi xel, NSA compression-tanneef-palanteer branches, hardware alignment ak scaling ak toppalante yu gudd.
Laaj 1 de 8
Lan moo waral seetlu sa bopp bu seer lool ci toppalante yu gudd?
Xoolaatal laaj bu nekk8 laaj
Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.
- Lan moo waral seetlu sa bopp bu seer lool ci toppalante yu gudd?
- Njëg li dafay màgg quadratically ak guddaayu toppalante bi
- Dafay jëfandikoo ay paramet yu néew
- Mënul dox ci GPU yi
- Dafay bàyyi xel ci token bu njëkk bi kese
- Lan mooy xalaat bu mag bi ci bàyyi xel ci blok-sparse?
- Di dindi feed-forward yi
- Liggéeyu xayma ci digganté ay bloku token yuñ tann rek
- Jëfandikoo benn boppu bàyyi xel kese
- Ñaari yoon jëmmal
- Lan moo wuutale NSA ak xeetu paj yu bari yu njëkk ya?
- Dafay dindi bépp bàyyi xel
- Du dox ludul ci diiru inference
- Dafay mëna tàggat ci njeexte ba ci njeexte gi te méngoo ak aparey bi
- Soxlawul benn GPU
- Ban ñatti bànxaas la NSA boole ngir am caabi ak valeur?
- Enkodeer, dekodeer, ak dugal
- Bàyyi, yamale ak boole
- Kantifikaasioŋ, dagg ak distilaasioŋ
- Kompresioŋ, tànneef, ak palanteer buy gliise
- Lan moo waral NSA di jëfandikoo ay operaasioŋ ci niveau block moo gën ñuy jëfandikoo ay sparsity ci niveau token?
- Ngir méngale ak jëfandikoo mémoire GPU ak génnug tensor-core ngir gaawaay dëgg
- Ngir gëna rëy model bi
- Ngir moytu jëfandikoo bàyyi xel
- Ngir wàññi limu diisaay yi
- Ci decodage autorégressif, bàyyi xel bu néew lu am solo la ndax jéego boobu dafay faral di nekk:
- calcul-lien ci kaw matrix yokk
- Memoire-bound, kon wàññi token yi daf koy gaaw
- Guddaayi fi ñuy toppalante mënul aju ci
- Dox ci CPU yi kese
- What information does NSA’s compression branch provide to the attention computation?
- A full dense score for every token pair
- Only the optimizer’s first moment
- Coarse-grained summaries of blocks of tokens
- A random subset of output labels
- Why can learned sparsity patterns be more useful than a fixed arbitrary mask?
- They remove the need to compute attention scores
- They guarantee zero memory use
- They force every token to attend globally
- They can select blocks relevant to the input while remaining aligned with efficient hardware operations