Questions about performance.

[Jacki1223]

Hi team! I found that for some benchmarks in /test, the performance of cutile python is much worse than that of Pytorch. Is this normal?

[ZhangZhiPku]

could you provide more specific benchmarks?

I tested cutile on a single RTX 5090, among gemm, flash attention, rmsnorm
cutile gemm is about 15%-50% slower than cutlass in common cases.
cutile flash attenion is as fast as the torch.sdpe.
cutile rmsnorm is 20% faster than PyTorch.