Tag
Nous Research introduces Lighthouse Attention, a training-only subquadratic wrapper for scaled dot-product attention that accelerates long-context pretraining and can be removed before deployment to preserve vanilla inference efficiency.