You raise the peak learning rate to reach the same pretraining loss in fewer steps, and a few thousand steps in the loss spikes and never recovers.

Before each spike the largest attention logits keep growing, and in several heads the softmax collapses until each query puts nearly all its weight on one key.