0.17.3 Release (#178)

Replaces 0.17.2, because 0.17.2 is broken on torch < 2.4. (We really
need automated testing.)

Fixes #177.

Fix python 3.8 and 3.9 releases (#129)

Fixes #128.

Release v0.17.0 (#125)

Fix attention autocast (#94)

Torch autocasts attention weights into FP32 because of softmax, but
doesn't autocast back into the user-specified data type.

Up until recently, we explicitly passed the autocast dtype in all
autograd function wrappers (reference:

https://github.com/SHI-Labs/NATTEN/blob/3b54c76185904f3cb59a49fff7bc044e4513d106/src/natten/functional.py#L149),
but this is wrong, because the user might be doing BF16.

According to the latest torch documentation, this has not been changed
since the first NATTEN release.

Because it's error prone, this commit explicitly calls cast on all
attention tensors to match the dtype of value. If it's already matching,
          torch will ignore it, and it shouldn't really get in the way
          of AMP mechanics.

Reference: #93

Update website (#83)

Co-authored-by: Ali Hassani <[email protected]>

0.14.6 Release

* Sync backend changes
  * Consistent C function signatures.

* Update changelog and readme

* Minor fix

PyTorch 2.0 Support (#21)

* PyTorch 2.0 wheels!

* Almost forgot

Merge pull request #4 from alihassanijr/main

Tiled 3x3 NA, torch 1.13 support, cp310 and cp311 wheels, and more.

Merge pull request #1 from alihassanijr/main

CPU kernels, refactored code, new release

Update URLs