Causal Masking

Masking technique preventing attention to future positions in autoregressive models.

Ensures token at position i can only attend to positions ≤ i. Essential for GPT-style models. Implemented as triangular mask.