@pama, if useful - here are utilization numbers for our attention backwards kernels (causal and non-causal, head dim = 64): https://github.com/HazyResearch/ThunderKittens/blob/main/att...