GLU Variants Improve Transformer

Proposed Gated Linear Unit (GLU) variants for the feed-forward layer of transformers, proving that SwiGLU consistently outperforms standard ReLU and GeLU activations.

Directly guides architectural decisions, alignment strategy, and serving infrastructure for task-text-generation.

Assumptions

• Empirical distribution regularity holds and target domain adheres to pretraining linguistic/visual support

Limitations

• Resource scaling, inference memory requirements, and alignment robustness vary with model size and hardware topology