[topwalktown]

Transformers like Llama use rotary embeddings which are applied in every single attention layer

https://github.com/huggingface/transformers/blob/222505c7e4d...

[markisus]

Very interesting! Do you know if there were any studies about whether this improves performance?