I have a transformer attention mechanism which seems to be more data-efficient than the usual dot product, and I'm trying to write a performant backwards kernel

I have a transformer attention mechanism which seems to be more data-efficient than the usual dot product, and I'm trying to write a performant backwards kernel for it.