Is it possible to support non-contiguous input tensor ?

[ProHuper]

The attention kernel implemented with ThunderKittens requires the sequence length to be a multiple of 128, and the input tensor to be contiguous. However, PyTorch's SDP kernel does not have these restrictions. If I pad and make the input tensor contiguous before feeding it, the speed of ThunderKittens will be much slower compared to PyTorch's SDP due to the introduction of significant memcpy operations.