Kedis ku xidhan hagaha · Adag Heerka

Kediska Feejignaanta Daahsoon ee Madax-badan

Sahaminta kaydka qarsoon ee MLA la wadaago, dib u dhiska darajada hoose, nuugista saadaalinta, dariiqa codaynta booska, iyo kaydinta xusuusta-xusuusta.

Wadooyinka haga ee la xidhiidhaFeejignaan Daahsoon oo Madax Badan
Su'aal 1 ee 8

Waa maxay dhibaatada aasaasiga ah ee Madax-bannaanida Daahsoon ee loogu talagalay in la yareeyo?

Horudhac su'aal kasta8 su'aalo

Jawaabaha iyo sharraxaadaha waxay soo baxayaan markaad qaadanayso kediska.

  1. Waa maxay dhibaatada aasaasiga ah ee Madax-bannaanida Daahsoon ee loogu talagalay in la yareeyo?
    • Qiimaha xusuusta ee kaydka-qiimaha-furaha inta lagu guda jiro falanqaynta
    • yaraanta xogta tababarka
    • Tirada qiyaasaha moodeelka
    • cabbirka erayada Tokenizer
  2. Sidee ayuu MLA u dhimaa kaydka KV?
    • Waxay tirtirtaa calaamadihii hore ee kaydka
    • Waxa ay kaydisaa vector yar oo dahsoon oo heer hoose ah halkii ay ka ahaan lahayd furayaasha iyo qiyamka madax kasta oo buuxa
    • Waxay u isticmaashaa miisaan 1-bit ah wax kasta
    • Waxay gabi ahaanba ka boodaa dareenka calaamadaha qaar
  3. Qaabkee ayaa markii ugu horreysay soo bandhigay Feejignaanta Daahsoon ee Madax-badan?
    • GPT-2
    • BERT
    • DeepSeek-V2
    • LLAMA 1
  4. Muxuu MLA ugu baahan yahay waddo 'kala go'ay' gaar ah oo loogu talagalay dhejinta booska wareega?
    • Sababtoo ah jagooyinku maaha kuwo ku habboon dareenka
    • Sababtoo ah RoPE ayaa labanlaabmaya ereyada
    • Sababtoo ah waxay dedejisaa optimizer-ka
    • Sababtoo ah wareegtada laguma laabi karo mashruuca kor u kaca ee la nuugay sida furaha intiisa kale
  5. Qiyaastii intee in le'eg ayay hoos u dhigtay xusuusta KV-cache ee DeepSeek-V2 ka soo warbixisay MLA iyo dareenka madax-madaxeed ee caadiga ah?
    • In ka badan 90%
    • Ilaa 5%
    • Dhab ahaan 50%
    • Wax dhimis ah ma jiro, kaliya guulaha xawaaraha
  6. Waa maxay faa'iidada la taaban karo ee si toos ah uga socotaa kaydka yar ee calaamad kasta?
    • Baabi'inta baahida GPU kasta
    • Cabbirrada kala-soocidda ee ka weyn iyo macnaha guud ee isla qalabkaas
    • Turjumaada tooska ah ee luuqadaha oo dhan
    • Ka saarida hawlgalka softmax
  7. What is the purpose of absorbing MLA up-projection weights into other projections?
    • Remove positional information from every token
    • Increase the size of the cached latent
    • Avoid materializing full per-head keys and values during inference
    • Replace attention with a recurrent network
  8. What positional information does MLA handle on a separate decoupled key path?
    • Class labels from the training data
    • Image segmentation maps
    • The output vocabulary probabilities
    • Rotary position embeddings that cannot be absorbed like the other key projections