引導式測驗 · 中等 等級

用於訓練測驗的現貨和搶佔式 GPU

使用可中斷的 GPU 容量進行可恢復訓練,同時考慮檢查點、提供者訊號、重新啟動開銷和有效成本。

問題 1 的 8

Why can spot or preemptible GPU capacity be interrupted?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. Why can spot or preemptible GPU capacity be interrupted?
    • The scheduler always replaces interrupted capacity without job recovery
    • The cloud instance automatically saves all training checkpoints
    • The provider may reclaim spare capacity for other demand
    • Interruptions occur only between completed training epochs
  2. Which checkpoint contents better support resuming training state?
    • Model weights without optimizer, progress, or random state
    • Model, optimizer, scheduler, progress, and relevant random state
    • A checkpoint that stores metrics but no training state
    • A checkpoint saved only on an instance that may be reclaimed
  3. What does checkpoint frequency trade off?
    • Checkpoint age and scheduler membership changes
    • Optimizer checkpoint size and training batch size
    • Restart count and accelerator memory capacity
    • Storage and pause overhead are traded against work lost after interruption
  4. Why should interruption notice timing not be generalized across clouds?
    • Signals and notice behavior depend on provider and resource type
    • All providers use identical interruption APIs and timing
    • A notice is guaranteed for every host or hardware failure
    • Checkpoint code controls whether the cloud sends a signal
  5. What can happen to other distributed workers when one rank is interrupted?
    • A missing worker can always be replaced without restarting other ranks
    • Other ranks can complete a collective with no matching participants
    • They may wait or fail at a collective unless the job coordinates recovery
    • Every process continues safely after any rank exits
  6. Which workload is a weaker fit for interruptible capacity?
    • An offline batch job with durable, tested checkpoints
    • A latency-critical endpoint with strict availability objectives
    • An embedding backfill that can resume by partition
    • Independent parameter-sweep trials with separate outputs
  7. Why compare cost per successful run rather than hourly rate alone?
    • The lowest hourly rate is the cost of a successful run
    • A GPU’s theoretical peak determines restart overhead
    • Cloud compute is free after an interruption
    • Restart, lost work, and checkpoint overhead change total cost
  8. What should be tested before relying on a recovery workflow?
    • A controlled interruption and checkpoint restore in staging
    • Restart the process without restoring any saved state
    • Use a different checkpoint format from production
    • Confirm only that a shutdown notice appears in logs