Add tuned triton MOE config for Qwen3-Omni

[sammysun0711]

Motivation

This PR aim to add tuned triton MOE config for Qwen3-Omni with TP=8 based on triton 3.4.0

Modifications

Accuracy Tests

Benchmarking and Profiling

Qwen3-Omni 8Kin/500out+10image (1280x960) TP=8
Before

============ Serving Benchmark Result ============
Backend:                                 sglang-oai-chat
Traffic request rate:                    inf
Max request concurrency:                 1
Successful requests:                     32
Benchmark duration (s):                  104.72
Total input tokens:                      523950
Total input text tokens:                 139310
Total input vision tokens:               384640
Total generated tokens:                  8014
Total generated tokens (retokenized):    6948
Request throughput (req/s):              0.31
Input token throughput (tok/s):          5003.24
Output token throughput (tok/s):         76.53
Total token throughput (tok/s):          5079.77
Concurrency:                             1.00
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   3271.41
Median E2E Latency (ms):                 3678.95
---------------Time to First Token----------------
Mean TTFT (ms):                          1242.87
Median TTFT (ms):                        1209.16
P99 TTFT (ms):                           1474.98
--------------Time per Output Token---------------
Mean TPOT (ms):                          8.11
Median TPOT (ms):                        8.09
P99 TPOT (ms):                           8.61
---------------Inter-Token Latency----------------
Mean ITL (ms):                           8.95
Median ITL (ms):                         7.80
P95 ITL (ms):                            15.68
P99 ITL (ms):                            23.62
Max ITL (ms):                            43.39
==================================================

After

============ Serving Benchmark Result ============
Backend:                                 sglang-oai-chat
Traffic request rate:                    inf
Max request concurrency:                 1
Successful requests:                     32
Benchmark duration (s):                  97.98
Total input tokens:                      524348
Total input text tokens:                 139708
Total input vision tokens:               384640
Total generated tokens:                  8014
Total generated tokens (retokenized):    6255
Request throughput (req/s):              0.33
Input token throughput (tok/s):          5351.57
Output token throughput (tok/s):         81.79
Total token throughput (tok/s):          5433.36
Concurrency:                             1.00
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   3061.08
Median E2E Latency (ms):                 3430.02
---------------Time to First Token----------------
Mean TTFT (ms):                          1172.24
Median TTFT (ms):                        1152.55
P99 TTFT (ms):                           1386.59
--------------Time per Output Token---------------
Mean TPOT (ms):                          7.56
Median TPOT (ms):                        7.56
P99 TPOT (ms):                           7.68
---------------Inter-Token Latency----------------
Mean ITL (ms):                           9.01
Median ITL (ms):                         7.60
P95 ITL (ms):                            15.23
P99 ITL (ms):                            22.69
Max ITL (ms):                            38.24
==================================================

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.