With such a high throughput because of sparsity, I'm particulary interested in distilling it into other architectures. I'd like to try a recurrent transformer when I have the time