reintroduce dropped na2d_qk and na2d_av APIs from 0.17.5

[tha-santacruz]

Hey there,

Some people at NVIDIA recently proposed a "Native Segmentation Vision Transformer" (https://arxiv.org/abs/2505.16993), where they say they use na2d_qk and na2d_av in their content-aware spatial grouping algorithm (see algorithm 2 in appendix E.1). Apparently it is quite essential to the practical use of their model, as shown in the runtime and memory analysis of appendix E.2

Since I saw in the changelog for release 0.20.0 that "unfused kernels may be revisited depending on demand and use case", I figured I would let you know that some people (like me) would be interested in this.

Appreciate all the great work — thanks!

[alihassanijr]

Thank you for your interest.

We'll look into bringing back those implementations (most likely new and better unfused implementations) in future releases as well.

[Bucunyear]

Hello, I wonder how can I use na2d_av in the latest version as a rookie instead of na2d?

[Bugjudger]

Me too

Hello, I wonder how can I use na2d_av in the latest version as a rookie instead of na2d?

[alihassanijr]

@Bucunyear @Bugjudger those APIs and their kernels have been removed since 0.20.0. We will try to bring some form of them back in future releases, but for now I'm afraid the only option is installing 0.17.5:

https://natten.org/install/#0175