Your gradient norm alert fires, but when you inspect the documents at the data cursor position you find nothing obviously wrong. What else could be causing the spike?

If the data appears clean, the issue is likely infrastructure or numerical. First, check for silent data corruption by comparing gradients across replicas. Second, verify the learning rate scheduler logs to ensure the LR value was correct. Third, check GPU memory logs for OOM events that might have triggered a silent fallback to lower precision. Finally, profile attention patterns on documents near the spike to identify numerical instability in specific attention heads triggered by long sequences.