add tutorial group gemm example

[scxiao]

Cherry pickup tutorial group gemm example

[zhanglx13]

I just tried it and here is what I got:

group-gemm-performance:
N cuBLAS Triton
0 128.0 0.061440 0.031040
1 256.0 0.035521 0.042240
2 512.0 0.060640 0.065440
3 1024.0 0.108961 0.392642

Does it look ok to you, @scxiao?

[scxiao]

I just tried it and here is what I got:

group-gemm-performance: N cuBLAS Triton 0 128.0 0.061440 0.031040 1 256.0 0.035521 0.042240 2 512.0 0.060640 0.065440 3 1024.0 0.108961 0.392642

Does it look ok to you, @scxiao?

Here is the a100 performance, unit is time (ms).
group-gemm-performance:
N cuBLAS Triton
0 128.0 0.020480 0.013312
1 256.0 0.023552 0.018432
2 512.0 0.032768 0.026624
3 1024.0 0.071680 0.087040

So, our performance is bad. Let me do some investigation, we should be much better than these numbers.

[scxiao]

Latest performance on MI250X
group-gemm-performance:
N cuBLAS Triton
0 128.0 0.061441 0.017120
1 256.0 0.034881 0.021281
2 512.0 0.060481 0.039040
3 1024.0 0.108481 0.153281

2x to 2.5x faster than first try

[scxiao]

Latest performance numbers on MI250
group-gemm-performance:
N cuBLAS Triton
0 128.0 0.061440 0.016480
1 256.0 0.035040 0.020160
2 512.0 0.060481 0.035040
3 1024.0 0.108640 0.113441