Failure Recovery on Large-scale Distributed Training
1Replay SizeJIT Checkpointing
Evaluation Results
| Method | Links | |
|---|---|---|
| JIT CheckpointingCheckpoint Timing=On failure, Recovery Mechanism=Restart + reconstruct state2026.07 | 1 | |
| DEADPOOLCheckpoint Timing=Per-step (in-memory), Recovery Mechanism=Node replacement + reconfiguration2026.07 | 1 |