引導式測驗 · 中等 等級

大規模容錯與 GPU 故障測驗

設計具有持久檢查點、運行狀況訊號、重啟策略和經過驗證的恢復行為的彈性分佈式訓練。

相關引導路徑大規模 GPU 故障容錯
問題 1 的 8

為什麼單一工作者故障會中斷分散式訓練任務?

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. 為什麼單一工作者故障會中斷分散式訓練任務?
    • Every rank can continue an update after any peer disappears
    • A process exit never affects other workers
    • Workers coordinate computation and synchronization across devices
    • Distributed updates require no communication between participants
  2. 忠實恢復訓練可能需要哪些狀態?
    • Model weights without optimizer or scheduler state
    • Model, optimizer, scheduler, progress, and relevant random state
    • Epoch count without data progress
    • Train-test split metadata without model state
  3. Where should recovery checkpoints be stored for node-loss recovery?
    • Ephemeral node disk
    • GPU memory
    • 主機頁面快取
    • Durable storage that survives the failed node
  4. 為什麼重新啟動策略應區分基礎設施故障和確定性錯誤?
    • Blind retries can repeat a code or data failure indefinitely
    • A transient timeout can never recover
    • Every process crash corrupts all training data
    • Restart repairs malformed input records
  5. 當彈性訓練調整工作者數量時,什麼可能會改變?
    • The number of data-loader workers only
    • The GPU model changes when workers are removed
    • Effective batch, data sharding, and training schedule
    • The parameter count changes with rank assignment
  6. 為什麼要在預演環境中終止工作者來測試恢復?
    • A staging run proves that hardware failures cannot recur
    • A controlled worker termination verifies checkpoint consistency and restart behavior
    • A restart test removes the need to keep durable checkpoints
    • A recovery test confirms convergence without restoring state
  7. What can GPU health monitoring reveal?
    • Class-specific prediction errors
    • Data split leakage
    • Optimizer convergence speed
    • Thermal, memory, power, driver, or interconnect issues
  8. What should be checked after a job resumes from a checkpoint?
    • Training progress and evaluation behavior remain consistent
    • A single rank restarted while others retained stale state
    • The checkpoint timestamp changed after loading
    • The worker count remains identical after a membership change