[3rdparty, document] Updated Documentation that for triton fused_moe kernel tuning for AMD Instinct GPUs

[kkHuang-amd]

Motivation

Updated Documentation for triton fused_moe kernel tuning for AMD Instinct GPUs.

Modifications

• Upload a tuning script file
• introduce the tuning parameters setting
• Provided example bash commands for tuning script run

Checklist

• [O] Update documentation as needed, including docstrings or example tutorials.

[ispobock]

Thanks for contributing this great tuning script! Could we use this script to tune on other devices? Any general steps for adaptation?

[kkHuang-amd]

Thanks for contributing this great tuning script! Could we use this script to tune on other devices? Any general steps for adaptation?

Yes, it can be used on other devices, but some parameter is the ROCm platform only, it needs to be removed when using on other platforms.

These two parameters only exist on ROCm

1. matrix_instr_nonkdim_range
2. kpack_range

[HaiShaw]

@ispobock Later we may come up with a generic script for more devices, after we collect device specific triton kargs extensions.
If this one merged first, we can modify the script pointing to the newer one later.

[zhyncs]

hold on
fused_moe_grok will be removed soon

[zhyncs]

ref #2223

[zhyncs]

ref #2225

[HaiShaw]

Thanks!

[zhyncs]

Thanks @kkHuang-amd @HaiShaw May you help update the doc from fused_moe_grok to fused_moe_triton in follow-up PR? Thanks!

[HaiShaw]

@zhyncs certainly, also notified @merrymercy for #2223