Ahmadzei's picture
added 3 more tables for large emb model
5fa1a76
- Combine the attention and causal masks into a single one, pre-computed for the whole model instead of every layer.