[diwank]

Seconded. I’m guessing you could create an implementation that is able to do that and then write optimised triton/cuda kernels to accelerate them but need to investigate further