Maswali ya Usambamba wa Data

Majibu na maelezo huonekana unapojibu maswali.

1. Katika ulinganifu wa kawaida wa data, kila GPU inashikilia nini?
   • Nakala kamili ya mfano, kusindika kipande tofauti cha kundi
   • Kipande tofauti cha tabaka za mfano
   • Hali ya kiboreshaji pekee, yenye uzani uliohifadhiwa katikati
   • Toleo lililobanwa la muundo na vigezo vichache
2. Ni operesheni gani ya mawasiliano huweka nakala za mfano katika kusawazisha kila hatua?
   • Matangazo, ambayo hunakili data moja ya GPU kwa wengine wote
   • Punguza yote, ambayo ni wastani wa gradient kwenye GPU zote
   • Tawanya, ambayo inagawanya data kati ya GPU
   • Hakuna mawasiliano yanayohitajika kati ya GPU
3. What memory limit remains in plain data parallelism?
   • The full model is sharded automatically across the GPUs
   • Only the input batch uses device memory
   • Each GPU keeps model parameters and optimizer state replicated, limiting the model size that fits
   • Each worker stores different model layers by default
4. Which factors does NCCL use when selecting collective algorithms?
   • Only the random seed and learning rate
   • Only the total number of model parameters
   • The optimizer type and validation score
   • Message size and topology
5. Je, PyTorch DistributedDataParallel inafichaje utulivu wa mawasiliano?
   • Kwa mwingiliano wa upinde rangi punguza kwa pasi inayoendelea ya kurudi nyuma
   • Kwa kuruka usawazishaji wa upinde rangi kwenye hatua zinazopishana
   • Kwa kubana gradients zote hadi nambari moja
   • Kwa kuendesha pasi ya mbele kwenye CPU
6. With equal local batch size B on each of W data-parallel workers, what is the global batch for one synchronized step?
   • B divided by W
   • B multiplied by W, before any gradient-accumulation factor
   • Always exactly W samples
   • The model parameter count divided by B
7. Why may adding more data-parallel GPUs fail to increase throughput proportionally?
   • All-reduce makes model parameters disappear
   • The model is copied only to one worker
   • Gradient communication, data loading, and synchronization add overhead
   • More workers always reduce the global batch to zero
8. What does gradient accumulation let a data-parallel training loop do?
   • Shard parameters automatically across all workers
   • Skip forward and backward computation
   • Replace the model with an all-reduce operation
   • Combine gradients from multiple microbatches before an optimizer update