Data Parallelism Quiz

Mhinduro uye tsananguro zvinoonekwa paunenge uchitora bvunzo.

1. Mune yakajairwa data parallelism, GPU yega yega inobata chii?
   • Ikopi yakakwana yemuenzaniso, kugadzirisa chidimbu chakasiyana chebatch
   • Chikamu chakasiyana chezvikamu zvemuenzaniso
   • Chete iyo optimizer state, ine huremu hwakachengetwa nechepakati
   • Iyo yakamanikidzwa vhezheni yemodhi ine mashoma paramita
2. Ndeipi mashandiro ekutaurirana anochengeta modhi replicas mukuwiriranisa nhanho imwe neimwe?
   • Broadcast, iyo inokopa data reGPU kune vamwe vese
   • Zvese-kudzikisa, izvo mavhareji magradients kune ese maGPU
   • Kuparadzira, iyo inokamura data pakati peGPUs
   • Hapana kutaurirana kunodiwa pakati peGPUs
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. PyTorch DistributedDataParallel inovanza sei kutaurirana latency?
   • Nekupindirana gradient zvese-kuderedza neinoenderera mberi yekumashure pass
   • Nekusvetuka kuwiriranisa kwegradient pamatanho anochinjana
   • Nekumanikidza ma gradients ese kune imwe nhamba
   • Nekumhanyisa kupfuudza kumberi paCPU
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