Maswali yanayounganishwa na mwongozo · Kati Kiwango

Maswali ya Kukagua Gradient

Inaelezea uundaji upya wa uanzishaji (activation rematerialization), akiba ya kumbukumbu, gharama ya kompyuta inayotegemea ratiba, na mchanganyiko na sharding na usahihi mchanganyiko.

Njia za mwongozo zinazohusianaUkagua gradient
Swali 1 ya 8

Je, ukaguzi wa gradient unafanya biashara gani ili kuokoa kumbukumbu?

Hakiki kila swali8 maswali

Majibu na maelezo huonekana unapojibu maswali.

  1. Je, ukaguzi wa gradient unafanya biashara gani ili kuokoa kumbukumbu?
    • Kokotoo ya ziada (recomputation) kwa matumizi ya chini ya kumbukumbu
    • Usahihi wa mfano kwa kasi
    • Nafasi ya diski kwa kipimo data cha mtandao
    • Usahihi kwa ukubwa wa kundi
  2. Kwa nini uanzishaji kawaida huhifadhiwa wakati wa kupita mbele?
    • Ili kuzionyesha kwa mtumiaji
    • Kwa sababu uenezaji wa nyuma unazihitaji ili kukokotoa viwango vya juu
    • Ili kuharakisha upakiaji wa data
    • Hazihitajiki kamwe
  3. In a simplified chain of N similarly sized layers, what can checkpoint boundaries spaced about sqrt(N) layers apart achieve in the textbook schedule?
    • It stores no activations and performs no recomputation
    • It guarantees constant memory for every model graph
    • Activation storage can scale on the order of sqrt(N) checkpoints
    • It makes optimizer-state memory scale as sqrt(N)
  4. What most determines gradient-checkpointing runtime overhead in a practical model?
    • Only the model’s file extension
    • A fixed 20–30% slowdown for every model
    • The number of labels in the dataset
    • Which operations are recomputed, checkpoint placement, graph structure, and hardware
  5. Kwenye PyTorch, ni matumizi gani ambayo hutumiwa kawaida kutumia ukaguzi wa gradient kwenye moduli?
    • sehemu.ya.matumizi.ya.kuhakiki
    • tochi.optim.SGD
    • tochi.nn.Kudondosha
    • tochi.cuda.empty_cache
  6. What happens to operations inside a checkpointed region during backward?
    • Their gradients are skipped entirely
    • Their forward computations are rerun to regenerate needed activations
    • Their weights are permanently discarded
    • They are executed only during inference
  7. Which training techniques can be combined with activation checkpointing to reduce different memory costs?
    • Only increasing the number of labels
    • Turning off backpropagation
    • FSDP for sharded model state and mixed precision for lower-precision tensors
    • Replacing the model with a tokenizer
  8. What is selective activation checkpointing intended to choose?
    • Which training labels are removed
    • Which model outputs are published
    • Which GPU becomes the tokenizer
    • Which operations to save and which to recompute based on their memory and runtime costs