0.0.7

numerically stable

0.0.6

no evidence that they used partial rotary embeddings, remove for furt…

0.0.5

forget about triton, optimize for clarity and education

0.0.4

remove cuda requirement for now

0.0.3

they use one-headed key/values in their transformer, fix

0.0.2

pre-layernorm for attention

0.0.1

update readme