Idanwo ti o sopọ mọ itọsọna · Lile Ipele

Iṣakojọpọ Ipinle Optimizer si Sipiyu ati NVMe Quiz

Apply optimizer offloading memory accounting, host update placement, and transfer tradeoffs, including why offloading does not remove all GPU allocations.

Jẹmọ itọsọna awọn ọnaOptimizer State Offloading To Sipiyu Ati Nvme
Ibeere 1 ti 8

Kini ibi-afẹde akọkọ ti gbigbe awọn ipinlẹ iṣapeju silẹ si Sipiyu tabi NVMe?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Kini ibi-afẹde akọkọ ti gbigbe awọn ipinlẹ iṣapeju silẹ si Sipiyu tabi NVMe?
    • Lati ṣe awoṣe diẹ sii deede
    • Lati dinku nọmba awọn akoko ikẹkọ ti o nilo
    • Lati yọkuro iwulo fun awọn gradients patapata
    • Lati baamu awọn awoṣe ti o tobi julọ nipa didi iranti GPU ti o ṣọwọn silẹ
  2. For an Adam setup with two FP32 moments and an FP32 master weight, how much memory do those three values use per parameter?
    • Two bytes for all three values
    • Twelve bytes, excluding model weights and gradients
    • Sixteen bytes for the three values alone
    • Zero bytes because they are implicit
  3. Ẹya ilana wo ni o gbajumọ ti iṣapeye iwọn-nla ti a gbejade?
    • Iṣẹ imuṣiṣẹ tuntun kan
    • JPEG aworan funmorawon
    • DeepSpeed ZeRO-Offload ati ZeRO-Infinity
    • Idinku gige
  4. Kini idiyele iṣẹ ṣiṣe akọkọ ti gbigbe si Sipiyu tabi NVMe?
    • Data gbọdọ rin irin-ajo lori awọn ọna asopọ ti o lọra bi PCIe tabi si disk, fifi lairi
    • Awoṣe gbagbe ikẹkọ iṣaaju
    • Optimizer di mathematiki ti ko tọ
    • GPU otutu silẹ ju kekere
  5. Bawo ni awọn ọna ṣiṣe piparẹ ṣe tọju pupọ ti lairi gbigbe naa?
    • Nipa fo sẹhin kọja
    • Nipa iṣaju data agbekọja pẹlu iṣiro GPU ti nlọ lọwọ
    • Nipa ikẹkọ nikan paramita kan ni akoko kan
    • Nipa piparẹ konge adalu
  6. Ni ZeRO-Offload, nibo ni igbesẹ Adam optimizer nigbagbogbo n ṣiṣẹ?
    • Ni kikun lori oludari NVMe
    • Lori a lọtọ àpapọ GPU
    • Lori Sipiyu, lilo awọn iwọn titunto si ati awọn akoko ti o fipamọ sibẹ
    • Inu awọn nẹtiwọki yipada
  7. Optimizer states are offloaded, but GPU weights and activations still exceed capacity. What follows?
    • The job can still run out of GPU memory
    • Offloading guarantees the job fits any GPU
    • The optimizer moments automatically become model activations
    • The model no longer needs forward tensors
  8. A CPU-offloaded optimizer becomes the training bottleneck. Which measurement is most relevant?
    • Only the inference accuracy
    • The UI font used for training logs
    • Only the number of dataset classes
    • Host optimizer-update time and host-device transfer time