Ikibazo gihuza ibibazo · Hagati Urwego

Slurm for AI Training Clusters Ikibazo

Kugenzura icyiciro cya Slurm, GPU no gukwirakwiza-gutangiza amategeko, guteganya ibice, imyitwarire yinyuma no kugenzura gukira.

Inzira zijyanye no kuyoboraSlurm Kubyiciro bya Training Training
Ikibazo 1 Bya 8

Ni irihe tegeko abakoresha basanzwe bakoresha kugirango batange akazi kamwe muri Slurm?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Ni irihe tegeko abakoresha basanzwe bakoresha kugirango batange akazi kamwe muri Slurm?
    • sbatch
    • kubectl
    • docker
    • git
  2. Nigute akazi ka Slurm gasaba GPUs?
    • Muguhindura inyandiko za DNS
    • Binyuze mu --gres = gpu
    • Binyuze kumurongo gusa
    • GPU ntishobora gusabwa muri Slurm
  3. Nibihe Slurm ifata ibyemezo byingengabihe?
    • slurmd
    • srun
    • slurmctld
    • sbatch
  4. Why can srun help launch distributed training tasks inside a Slurm allocation?
    • It automatically restores every model checkpoint
    • It replaces slurmctld as the scheduler
    • It converts a single GPU into multiple physical GPUs
    • It launches coordinated tasks across allocated nodes; the training setup may still need explicit DDP configuration
  5. Niyihe ntego yo gusubira inyuma muri Slurm?
    • Gukora imirimo migufi mu cyuho kidatinze imirimo idahwitse
    • Gusiba imirimo ishaje
    • Kurenga GPUs
    • Kugirango uhishe inyandiko zakazi
  6. What must an AI training script do so a requeued Slurm job can continue after a node failure?
    • Assume Slurm stores model weights automatically
    • Save checkpoints and reload them when the job restarts
    • Disable resource accounting
    • Submit a new job without restoring state
  7. Which Slurm daemon normally runs on each compute node to launch tasks and report node status?
    • slurmctld
    • sbatchd
    • slurmd
    • gresctl
  8. What can Slurm resource accounting help cluster administrators track?
    • The model’s validation accuracy automatically
    • The semantic content of training data
    • Whether every checkpoint is scientifically valid
    • Job resource use such as allocated CPUs, GPUs and time