Gemm benchmark for #3290: replaced torch._scaled_mm with torch.nn.functional.scaled_mm

[agolajko]

Summary

As discussed with @vkuzo in #3290

Replaced torch._scaled_mm with torch.nn.functional.scaled_mm and ran the two benchmark (bench_1x128_128x1_gemms.py and bench_1x128_128x128_gemms.py) scripts from here

Results on an H100 with the following setup:

Torchao: 0.15.0+git1fbc5f6a5
Python: 3.12.3 (main, Aug 14 2025, 17:47:21) [GCC 13.3.0]
PyTorch: 2.10.0.dev20251113+cu129
CUDA: 12.9
CuDNN: 91002

[OS]
OS: Linux 6.8.0-60-generic
Distribution: Ubuntu 24.04.3 LTS
570.133.20, NVIDIA H100 PCIe, 9.0

# python bench_1x128_128x1_gemms.py
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:02<00:00, 31.40s/it]
    M     N     K  out_dtype         bf16_mm_us    fp8_triton_us    fp8_scaled_mm_us    bf16 tflops/sec    triton tflops/sec    scaled_mm tflops/sec
-----  ----  ----  --------------  ------------  ---------------  ------------------  -----------------  -------------------  ----------------------
16640  5120  8192  torch.bfloat16       3223.73          4511.23             2405.09            432.997              309.42                  580.38
16640  8192  5120  torch.bfloat16       3243.3           4708.93             2404.93            430.385              296.429                 580.418

# python bench_1x128_128x128_gemms.py
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:16<00:00,  8.44s/it]
    M     N     K  out_dtype         bf16_mm_us    fp8_triton_us    fp8_scaled_mm_us    bf16 tflops/sec    triton tflops/sec    scaled_mm tflops/sec
-----  ----  ----  --------------  ------------  ---------------  ------------------  -----------------  -------------------  ----------------------
16640  5120  8192  torch.bfloat16       3351.36          4665.82             2170.48            416.507              299.168                 643.113
16640  8192  5120  torch.bfloat16       3466.82          4681.14             2286.5             402.636              298.189                 610.482

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3342

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[agolajko]

@vkuzo lmk if you have comments or suggestions for the above changes