[QST] How to use Grouped_gemm when MNK is not a regular shape

[chenhongyu2048]

What is your question?

Hi all, I'm having some issues with the 57_cutlass_grouped_gemm.
I need to do some gemm calculations where there may be a matrix form like 1024xNx1024, but N may not be a multiple of 8.
I believe this is common, but how should I modify my code to accommodate this kind of problem?

[chenhongyu2048]

Supplement:
I tried tweaking alignment, but it went against the needs of the TMA.

[thakkarV]

You'll have to either pad your allocations to be 16B aligned or use the cpasync based grouped GEMM kernel instead

[chenhongyu2048]

Thanks, I'll try it.