0.0.23

first take care of the block diagonal causal in fine attention

0.0.22

make a decision to deviate from their diagram, where the last token o…

0.0.21

start accumulating some different compression network ideas

0.0.20

just improvise a solution for compress and selection block sizes not …

0.0.19

make sure deepseek proposal can be compared to attention with gqa

0.0.18]

Full Changelog: 0.0.17...0.0.18]

0.0.18

Full Changelog: 0.0.17...0.0.18

0.0.17

Full Changelog: 0.0.16...0.0.17

0.0.16

Full Changelog: 0.0.15...0.0.16

0.0.15

handle rotary embeddings for sliding windows explicitly