Is cutlass hopper warp specialization gemm requires M alignment?

[MARD1NO]

Recently I want to run M=2 K=8192 N=7168 54_hopper_fp8_warp_specialized_gemm, it raise a cutlass error:

Got cutlass error: Invalid status at: 488

but in M=16/32 it works well, is there exist some configurations I need to change to support M=2 gemm？

[thakkarV]

Is your layout M major? If so, it has to respect TMA alignment.

[MARD1NO]

I just use example's layout, A is RowMajor, B is ColMajor, C is ColMajor（I think it can get the best performance since fp8 tensorcore requires this layout?), and I'm not very clear about what you said M Major

[MARD1NO]

emmm I change C -> RowMajor it works