[BUG]

Describe the bug
split_k_slices option does not appear to do anything on the profiler.

Steps/Code to reproduce bug

./cutlass_profiler --operation=Gemm --m=256 --n=256 --k=8192 --A=f16:column --B=f16:row --C=f16:column --split_k_slices=1,2,4,8,16

Expected behavior
For such a skinny GEMM, split_k should have a significant effect on the performance. But there is not difference in performance between the different splits.

Environment details (please complete the following information):

• Bare-metal, H100 PCIe
• CUTLASS 3.6.0

Additional context
Both at 1 and 16 splits the profiler measured ~26900 GFLOP/s on our device. The best performing ones for the two splits were:

# 1 split
1,CUTLASS,gemm,cutlass3x_sm90_sptensorop_s64x128x32spgemm_f16_f16_f32_void_f16_128x128x64_1x2x1_0_ntn_align8_warpspecialized_cooperative_epi_nosmem,passed,success,universal,256,256,8192,f16:column,f16:row,f16:column,f16:column,1,0,serial,1,1,heuristic,1,sptensorop,f32,128,128,64,1,2,1,9,4,2,1,64,128,32,90,90,8519680,1073872896,126,0.0398454,199.134,26951
# 16 splits
5,CUTLASS,gemm,cutlass3x_sm90_sptensorop_s64x128x32spgemm_f16_f16_f32_void_f16_128x128x64_1x2x1_0_ntn_align8_warpspecialized_cooperative_epi_nosmem,passed,success,universal,256,256,8192,f16:column,f16:row,f16:column,f16:column,1,0,serial,16,1,heuristic,1,sptensorop,f32,128,128,64,1,2,1,9,4,2,1,64,128,32,90,90,8519680,1073872896,126,0.039881,198.956,26927

We also tried less extreme shapes, but they all seemed to disregards split_k_slices.

Reproduce the report with the provided cutlass_profiler GEMM command on the stated H100 and compare results for split_k_slices values from 1 through 16. Trace how the profiler passes and reports split_k_slices, then verify that the selected configurations or measured performance reflect the requested splits rather than remaining identical.