Can NATTEN be integrated to TensorRT？

[QishuoLu]

No description provided.

[alihassanijr]

Thank you for your interest.

NATTEN kernels are almost entirely independent of PyTorch, and the backend API is specifically designed to avoid using torch's tensor API as to make binding to other frameworks/engines easier.

However, my understanding is that there is "partial" support for binding custom ops in TRT, and it's only limited to inference, so unless a model that depends on NATTEN is going to be served with TRT, I think it should be possible.