v0.2.2

Build cuda12 version with Hopper support.

v0.2.1

Bump minor version to 0.2.1.

v0.2.0

Expanded vmap support for flash_mha. Vmapping q but not k,v reduces t…

…o a grouped-query attention, which we now support.

v0.1.0a3

Try cibuildwheel.

v0.1.0a2

Try cibuildwheel.

v0.1.0a1

Set CUDA_HOME for the sdist

v0.1.0

Try release with github actions and rebased repo.

v2.5.5a2

Implement ring attention backward pass. More tests.

v2.5.5a1

Merge up to v2.5.5

v2.5.0a4

Implement custom sharding for flash_mha, to allow efficient multi-gpu…

… computation when sharded across batch or head dimensions.