5fa1a76
1
- Combine the attention and causal masks into a single one, pre-computed for the whole model instead of every layer.
