0.5.20

fix maximum update in triton flash attn

0.5.19

What's Changed

• wip zigzag by @lucidrains in #22

Full Changelog: 0.5.18...0.5.19

0.5.18

address https://github.com/lucidrains/ring-attention-pytorch/issues/21

0.5.17

Full Changelog: 0.5.10...0.5.17

0.5.12

Full Changelog: 0.5.10...0.5.12

0.5.10

keep numerator and denominator in float32 for tree attn decoding

0.5.9

some asserts and cleanup for tree attn decoding

0.5.8

make tree attention decoding work with triton flash attention forward

0.5.6

account for scenario where keys and values are already sharded in tre…

0.5.5

Full Changelog: 0.5.4...0.5.5