Kedis ku xidhan hagaha · Adag Heerka

U-qaadista Gobolka Optimizer ee CPU iyo Kediska NVMe

Codso optimizer offloading memory accounting, meelaynta cusboonaysiinta martida loo yahay, iyo wareejinta ganacsiyada, oo ay ku jirto sababta ka qaadistu aanay meesha uga saarin dhammaan qoondaynta GPU.

Wadooyinka haga ee la xidhiidhaSoo-dejinta Gobolka Optimizer ee Kumbuyuutarka iyo Nvme
Su'aal 1 ee 8

Waa maxay himilada koowaad ee u dajinta dawladaha kor u qaadaya CPU ama NVMe?

Horudhac su'aal kasta8 su'aalo

Jawaabaha iyo sharraxaadaha waxay soo baxayaan markaad qaadanayso kediska.

  1. Waa maxay himilada koowaad ee u dajinta dawladaha kor u qaadaya CPU ama NVMe?
    • Si loo sameeyo qaabka mid sax ah
    • Si loo yareeyo tirada xilliyada tababarka ee loo baahan yahay
    • Si loo baabi'iyo baahida loo qabo gradients gabi ahaanba
    • Si loo waafajiyo moodooyinka waaweyn iyadoo la xoreynayo xusuusta GPU-da yar
  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. Qaabkee ayaa sifeynaya soo dejin heersare ah oo caan baxay?
    • Hawl firfircooni cusub
    • Isku-buufinta sawirka JPEG
    • DeepSpeed ​​​​Zero-Offload iyo ZeRO-Infinity
    • Goynta si tartiib ah
  4. Waa maxay qiimaha waxqabadka ugu weyn ee u dejinta CPU ama NVMe?
    • Xogtu waa in ay ku socotaa xiriiryo gaabis ah sida PCIe ama diskka, ku dar daahida
    • Qaabka ayaa ilaaway tababarkii hore
    • Hagaajinta waxay noqotaa xisaab ahaan khalad
    • Heerkulka GPU ayaa aad hoos ugu dhacay
  5. Sidee bay nidaamyada rarista u qariyaan inta badan daahitaanka wareejinta?
    • Adigoo ka boodaya baaska danbe
    • Iyadoo la isku duba ridayo xogta horudhac u ah xisaabinta GPU socota
    • Adigoo tababaraya halbeeg keliya markiiba
    • Adigoo curyaaminaya saxnaanta isku dhafan
  6. Gudaha ZeRO-Offload, halkee ayay talaabada hagaajinta Adam inta badan ka fulisaa?
    • Gebi ahaanba kontaroolaha NVMe
    • Muuqaal gooni ah GPU
    • Kumbuyuutarka, adoo isticmaalaya miisaanka sayidkiisa iyo daqiiqadaha halkaas lagu kaydiyay
    • Gudaha dareemaha shabakada
  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