Guide-yakabatana mibvunzo · Zvakaoma Nhanho

Multi-Head Latent Attention Quiz

Ongorora MLA's yakagovaniswa cache, yakaderera-chinzvimbo kuvakazve, fungidziro yekunyudza, chinzvimbo-encoding nzira, uye inference-memory kuchengetedza.

Nzira dzekutungamira dzinoenderanaMulti Head Latent Attention
Mubvunzo 1 ye 8

Nderipi dambudziko rekutanga Multi-Head Latent Attention rakagadzirirwa kuderedza?

Tarisa mubvunzo wese8 mibvunzo

Mhinduro uye tsananguro zvinoonekwa paunenge uchitora bvunzo.

  1. Nderipi dambudziko rekutanga Multi-Head Latent Attention rakagadzirirwa kuderedza?
    • Mutengo wekurangarira wekiyi-value cache panguva yekufungidzira
    • Kushaikwa kwedata rekudzidzisa
    • Nhamba yemuenzaniso parameters
    • Tokenizer saizi yemazwi
  2. MLA inoderedza sei cache yeKV?
    • Iyo inobvisa ma tokeni ekare kubva kune cache
    • Inochengeta diki-rank-latent vector pane yakazara-yemusoro makiyi uye kukosha
    • Inoshandisa 1-bit huremu kune zvese
    • Inodarika kutarisa zvachose pane mamwe ma tokens
  3. Ndeipi modhi yakatanga kuunza Multi-Head Latent Attention?
    • GPT-2
    • BERT
    • DeepSeek-V2
    • LLaMA 1
  4. Sei MLA ichida nzira yakaparadzana 'yakaparadzaniswa' yekumisikidzwa kwenzvimbo inotenderera?
    • Nekuti zvinzvimbo hazvinei nekutariswa
    • Nekuti RoPE inopeta kaviri mazwi
    • Nekuti inomhanyisa optimizer
    • Nekuti kutenderera hakugone kupetwa mune yakanyudzwa kumusoro-fungidziro senge yasara kiyi inogona
  5. Zvichida kuti yakawanda sei KV-cache ndangariro kuderedza iyo DeepSeek-V2 yakashuma kubva kuMLA inopesana neyakajairwa-yepamusoro-soro kutarisa?
    • Kupfuura 90%
    • Zvingangoita 5%
    • 50% chaiyo
    • Hapana kudzikiswa, kukurumidza kuwana chete
  6. Ndeipi bhenefiti inoshanda inotevera zvakananga kubva kuMLA diki pa-token cache?
    • Kubvisa kudiwa kwechero GPU
    • Yakakura inference batch saizi uye kureba mamiriro pane imwechete hardware
    • Shandura yega yega mumitauro yese
    • Kubviswa kwekushanda kweSoftmax
  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