[Perf] fused_moe: add int4_w4a16 benchmark support and tuning config

[mgehre-amd]

Add int4_w4a16 quantization dtype to the MoE benchmark/tuning script with proper uint8-packed weight generation and group-wise scales.

The group_size is automatically extracted from the model config, supporting AWQ/GPTQ style (direct 'group_size' key) and compressed-tensors style (nested in 'config_groups').

Add tuned Triton MoE kernel configurations for int4_w4a16 quantization on the AMD Radeon 8060S Graphics (Strix Halo, gfx1151) for batch sizes 1-32. Model shape: E=128, N=768 (e.g. Qwen3-30B-A3B MoE layers).

Benchmarked on AMD Strix Halo (gfx1151, LPDDR5X-8000 128 GB):

Model: RedHatAI/Qwen3-30B-A3B-Instruct-2507.w4a16
input-len=128, output-len=128, num-prompts=5:
Median TTFT: 182.17 ms -> 141.69 ms (22% faster)
Median TPOT: 41.88 ms -> 39.73 ms (5% faster)

[gemini-code-assist]

Code Review

This pull request introduces support for int4_w4a16 quantization in the Fused MoE benchmark and tuning script, which is a great addition for performance evaluation on newer hardware. The changes are well-structured, including proper weight generation, scale handling, and a new utility to extract group_size from model configurations. The added tuning configuration for AMD Radeon 8060S Graphics is also valuable. I have one suggestion to improve the robustness of the new get_quantization_group_size function to prevent potential crashes from malformed configuration files.

[mergify]

Hi @mgehre-amd, the pre-commit checks have failed. Please run:

uv pip install pre-commit
pre-commit install
pre-commit run --all-files

Then, commit the changes and push to your branch.

For future commits, pre-commit will run automatically on changed files before each commit.

Tip

Is mypy or markdownlint failing?

mypy and markdownlint are run differently in CI. If the failure is related to either of these checks, please use the following commands to run them locally:

# For mypy (substitute "3.10" with the failing version if needed)
pre-commit run --hook-stage manual mypy-3.10
# For markdownlint
pre-commit run --hook-stage manual markdownlint

[tjtanaa]

LGTM