Releases: bdashore3/flash-attention
v2.6.3
Synced to Upstream version
NOTE: Backward and Dropout are disabled meaning that this release is INFERENCE ONLY.
This is because including these features more than doubles the build time and makes the github action time itself out. Please raise an issue to the parent repo to help reduce the build times if you want these features.
v2.6.1
Actions: Switch to CUDA 12.3 Signed-off-by: kingbri <[email protected]>
v2.5.9.post2
Quick release to add softcapping commits. Does not have backward, dropout, or alibi support.
v2.5.9.post1
Actions: Clarify dispatch formatting Signed-off-by: kingbri <[email protected]>
v2.5.8
Same as Upstream tag
Now built for only torch 2.2.2 and 2.3.0
v2.5.6
Same as Upstream tag
v2.5.2
Same as the upstream tag
Adds this PR to help fix building on Windows
v2.4.2
Inline with the parent repo's tag
Made for cuda 12.x and pytorch 2.1.2 and 2.2
v2.4.3 and up cannot be built on Windows at this time.
v2.4.1
Add Windows workflows
2.3.3-windows
In parity with the original tag
Built with Pytorch 2.1.1 and CUDA 12.2. This wheel will work with pytorch 2.1+ and CUDA 12+
Full Changelog: https://github.com/bdashore3/flash-attention/commits/2.3.3