[ROCm] Enable VLLM triton FP8 moe for gfx1201, tuned for Qwen3-30B-A3B-FP8 tp=2

[big-yellow-duck]

Purpose

gfx12xx cards support FP8, so we enabled FP8 Triton MoE in VLLM and tuned for Qwen/Qwen3-30B-A3B-Instruct-2507-FP8.

Test Plan

benchmark Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 with triton moe tuned on 2 Radeon PRO 9700

VLLM_ROCM_USE_AITER=0 vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 -tp 2 --enable-expert-parallel

Test Results

Benchmark using Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 on 2x Radeon PRO 9700

TTFT (ms)

ISL-OSL	Triton MoE Tuned
512-512	982.67
1024-1024	711.35
2048-2048	1250.83
4096-4096	4329.35
8192-1024	11351.34
16384-2048	131106.68
Average	24955.37

TPOT (ms)

ISL-OSL	Triton MoE Tuned
512-512	33.18
1024-1024	34.94
2048-2048	36.65
4096-4096	45.05
8192-1024	73.46
16384-2048	70.46
Average	48.95

E2E Latency (ms)

ISL-OSL	Triton MoE Tuned
512-512	17936.15
1024-1024	36449.97
2048-2048	76264.82
4096-4096	188823.78
8192-1024	86497.21
16384-2048	275340.95
Average	113552.15

Accuracy checks

GSM8K Accuracy

Metric	Triton MoE Tuned
exact_match,strict-match	83.40%
exact_match,flexible-extract	86.43%

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[tjtanaa]

This should not be include here because this is only about pushing the triton tuned config json