custom cuda kernel for {2, 3}d relative attention with pytorch wrapper - View it on GitHub