A transformer efficiency method that reduces context memory use by up to 8x and doubles inference speed without accuracy loss.

A faster implementation of normalization methods like RMSNorm, LayerNorm, and Dynamic Tanh used in large language models.