quiz bu lëkkale ak tegtal · Diggante Niveau

Aktiwisaasioŋ Rekomputaasioŋ Quiz

Test aktivaasioŋ biy saytu njariñu mémoire bi, xaymawaat buñ tànn, ak jubbantiwaat, boole ci lu tax xayma bu gëna bari du saafara bépp njëgu mémoire.

Yooni njiit yu jëm ciAktiwisaasioŋ xaymawaat kompromis
Laaj 1 de 8

Lan la recomputation biy aktive ngir sakkanal mémoire bi?

Xoolaatal laaj bu nekk8 laaj

Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.

  1. Lan la recomputation biy aktive ngir sakkanal mémoire bi?
    • model bu jaar yoon
    • Yokkateg ordinatër (FLOPs)
    • Yaatuwaayu band reso bi
    • Njëg ci lim
  2. Lan moo waral ñuy denc aktivaasioŋ forward-pass yi ci anam wu jaar yoon?
    • Ngir wane leen jëfandikukat bi
    • Ngir kompresse model bi
    • Dañu leen wara jëfandikoo ngir tokenisation
    • Ndax tasaare ci ginaaw daf leen soxla ñu xayma degrade yi
  3. Why should activation checkpointing overhead be measured on the actual model?
    • Checkpoint placement, operators, and hardware determine the cost of replay
    • It always adds exactly forty percent to wall-clock time
    • It guarantees a tenfold speed improvement
    • It removes all forward computation
  4. For an idealized N-layer chain with comparable activations, what storage scaling can square-root checkpoint placement achieve?
    • O(N squared) for all model parameters
    • O(1) for all training memory regardless of N
    • O(sqrt(N)) for activation storage
    • O(N squared) for optimizer state
  5. Lan mooy xalaat bi ci ginaaw tànneef (bu matwul)
    • Xaymaat bépp jëf ñaari yoon
    • Bul musa xayma dara
    • xaymawaat couche biy dugal rek
    • Recompute rek memory bu diis waaye ops yu yomb, cache yu seer yi
  6. Ban pexem mottali lañuy faral di boole ak xaymawaat ngir sakkanal memory bu gëna bari?
    • Yokk dayo lots bi
    • Yebbi mbir yi ci CPU wala NVMe
    • Dindi bépp diisaayu normalisasioŋ
    • Wecci ci mode inference
  7. A checkpointed block contains dropout. Why does consistent RNG state matter during replay?
    • Dropout becomes a model-weight quantizer
    • The optimizer no longer needs gradients
    • Backward recomputation should reproduce the intended forward randomness
    • The batch size must become one
  8. Checkpointing reduces stored activations, but optimizer states still exceed GPU memory. Why is the job still out of memory?
    • Activation recomputation does not remove optimizer-state storage
    • Checkpointing increases every optimizer tensor to FP64
    • Backward no longer consumes any memory
    • The model cannot use checkpointing with an optimizer