1-Bit iyo Ternary BitNet Model Kedis

Jawaabaha iyo sharraxaadaha waxay soo baxayaan markaad qaadanayso kediska.

1. Waa maxay sababta BitNet b1.58 loogu sifeeyay isticmaalka ku dhawaad ​​1.58 bits miisaankiiba?
   • Sababtoo ah miisaan kastaa wuxuu leeyahay saddex qiime oo suurtagal ah (-1, 0, +1) iyo log2(3) waa qiyaastii 1.58
   • Sababtoo ah waxay ku wareegsan tahay 2 bits
   • Sababtoo ah waxay isticmaashaa 1.58 GB oo xusuusta ah
   • Sababtoo ah 1.58 waa celceliska 1 iyo 2 bits
2. Maxaa ka dhigaya hawlaha matrixka BitNet ka jaban yihiin moodooyinka caadiga ah?
   • Waxay ka boodaan habka dareenka
   • Miisaanka -1, 0, +1 waxay isu beddelaan isku dhufashada isu-geynta iyo kala-goynta
   • Waxay isticmaalaan erayo ka yar
   • Waxay ku shaqeeyaan oo keliya TPU-yada
3. Farqiga muhiimka ah ee u dhexeeya BitNet iyo qiyaasta tababarka ka dib ee caadiga ah waa BitNet:
   • Kaliya qiyaaso ka dib marka tababarka la dhammeeyo
   • Waxba lama tababari karo
   • Waxaa laga soo tababaray meel eber ah oo miisaan yar leh
   • Waxay isticmaashaa miisaan 32-bit gudaha ah
4. Inta lagu jiro tababarka, sidee ayay BitNet u cusboonaysiisaa miisaanka inkasta oo aan la kala saarin tillaabada qiyaasidda?
   • Waxay dejisaa dhammaan miisaanka
   • Waxay ku tababartaa gradients la'aan
   • Waxa ay isticmaashaa barashada xoojinta kaliya
   • Waxay isticmaashaa qiyaase toosan oo leh miisaanyo hadh-sare oo sax ah
5. Waa maxay faa'iidada dheeraadka ah ee u oggolaanaysa qiimaha 0 (ternary, not binary saafi ah) ayaa bixiya?
   • Ku-dhisnaanta yar-yar, maadaama miisaanka eber la dhaafi karo
   • Saxnimo ka sarraysa barta sabaynaysa
   • Isku xidhka intarneedka oo dhakhso badan
   • Cabbirka ereyada ka weyn
6. Waa maxay caqabada ugu weyn ee qalabka si loo xaqiijiyo guulaha waxtarka buuxa ee BitNet?
   • Ma jiraan GPU-yo ka hadhay adduunka
   • Inta badan jajabyada hadda jira waxaa loo hagaajiyay xisaabta dul sabaynaysa, maaha kuwa lagu daro-keliya ternary ops
   • Moodooyinka ternary waxay u baahan yihiin kombuyuutarrada quantum
   • Moodooyinka aad bay u yar yihiin arrintu
7. If a BitNet weight is zero, what does it contribute to a dot product?
   • It multiplies the activation by negative one
   • It adds a full-precision matrix multiply
   • It contributes zero and can potentially be skipped by suitable hardware
   • It changes the number of model layers
8. Why are higher-precision shadow weights retained during BitNet training?
   • To perform inference with full-precision weights only
   • To store the tokenizer vocabulary
   • To replace the straight-through estimator with exact gradients
   • To maintain trainable values for gradient updates despite low-bit forward quantization