What is attention collapse in transformer-based Keras models?

Attention collapse occurs when attention weights become overly concentrated or uniform.

In deep transformer stacks, attention heads may stop learning meaningful distributions and collapse into trivial patterns. This reduces representational diversity and model performance.