Ikibazo gihuza ibibazo · Biragoye Urwego

Urupapuro rwitonderwa hamwe na vLLM Ikibazo

Ikizamini CyapapuroKwitondera KV guhagarika, gushushanya-kumubiri-gushushanya, no kugabana, harimo gutandukanya ububiko bwa cache yibitse hamwe nububiko bwikitegererezo.

Inzira zijyanye no kuyoboraUrupapuro rwitondewe na Vllm
Ikibazo 1 Bya 8

Niki 'KV cache' ibika mugihe cyo kwandika inyandiko?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Niki 'KV cache' ibika mugihe cyo kwandika inyandiko?
    • Icyitegererezo cyamahugurwa
    • Ibyangombwa byinjira byumukoresha
    • Urufunguzo nagaciro kerekana ibimenyetso byashize kugirango ibimenyetso bishya bishobora kubitabira
    • Kopi ikomatanyije yuburemere bwikitegererezo
  2. Ni ubuhe buryo bwa sisitemu y'imikorere yahumekeye Paged?
    • Ibikoresho byo gukuraho disiki
    • Paji yibuka
    • Guhagarika dosiye
    • Guhanura ishami rya CPU
  3. Ni ikihe kibazo kijyanye no kugabura kwa KV gakondo PagedIbyifuzo bikemura?
    • Buhoro buhoro umurongo wa interineti
    • Kubura amakuru yamahugurwa
    • Intege nke z'umutekano wibanga
    • Imyanda yibuka kubitsa ibinini binini bihuza buri cyifuzo
  4. Nigute PagedIbyitonderwa bireka ibisubizo byinshi bigabana kwibuka kubintu bisanzwe?
    • Urutonde rutandukanye rwerekana page imwe yumubiri ukoresheje kopi-ku-kwandika
    • Isiba ikibazo nyuma yo gukoresha bwa mbere
    • Irabika ikibazo kuri disiki
    • Isubiramo icyitegererezo kuri buri gisohoka
  5. VLLM na PagedIbyifuzo byatejwe imbere he?
    • Laboratwari yo mu myaka ya za 90
    • UC Berkeley (yasohowe 2023)
    • Itsinda ryamamaza ibicuruzwa
    • Ihuriro ritazwi kumurongo
  6. Ni izihe nyungu nyamukuru za vLLM kumuntu ukorera LLM?
    • Ingero nini yerekana neza ibipimo
    • Ikimenyetso cyikora
    • Ibyuma bya GPU kubuntu
    • Byinshi byinjira kandi bisabwa byinshi kuri GPU
  7. A sequence’s logical KV block 3 resides in a different physical location. What tells PagedAttention where to read it?
    • The sequence’s block table mapping logical blocks to physical blocks
    • The model’s class-label list
    • The learning-rate scheduler
    • A requirement that every block follow block 2 contiguously
  8. A PagedAttention cache ends with a half-filled block. Which statement about memory waste is accurate?
    • Paging guarantees every allocated byte contains a token
    • The entire model must be copied for each unused slot
    • Unused slots can remain in that final block despite reduced allocation fragmentation
    • The block table cannot represent a partial sequence