Add fused multimodal ROPE RMS for Qwen vision language models

[xytpai]

Motivation

Add mrope fused kernel for Qwen-vl models with inference mode.

Technical Details

For norm implementation details, we use warp norm method for better performance, overall uplift 10x compared with native torch ops.

Tests

mrope qknorm fused 308 test (Qwen3-VL-235B-A22B-Instruct-FP8-dynamic)

before:

============ Serving Benchmark Result ============
Backend:                                 sglang-oai-chat
Traffic request rate:                    inf       
Max request concurrency:                 64        
Successful requests:                     128       
Benchmark duration (s):                  119.44    
Total input tokens:                      148715    
Total input text tokens:                 68459     
Total input vision tokens:               80256     
Total generated tokens:                  129952    
Total generated tokens (retokenized):    77974     
Request throughput (req/s):              1.07      
Input token throughput (tok/s):          1245.07   
Output token throughput (tok/s):         1087.98   
Total token throughput (tok/s):          2333.05   
Concurrency:                             48.37     
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   45137.45  
Median E2E Latency (ms):                 45453.06  
---------------Time to First Token----------------
Mean TTFT (ms):                          3902.20   
Median TTFT (ms):                        1605.80   
P99 TTFT (ms):                           9031.68   
--------------Time per Output Token---------------
Mean TPOT (ms):                          42.72     
Median TPOT (ms):                        42.93     
P99 TPOT (ms):                           76.06     
---------------Inter-Token Latency----------------
Mean ITL (ms):                           53.58     
Median ITL (ms):                         37.15     
P95 ITL (ms):                            111.88    
P99 ITL (ms):                            220.84    
Max ITL (ms):                            6631.77   
==================================================

after:

============ Serving Benchmark Result ============
Backend:                                 sglang-oai-chat
Traffic request rate:                    inf       
Max request concurrency:                 64        
Successful requests:                     128       
Benchmark duration (s):                  111.89    
Total input tokens:                      148823    
Total input text tokens:                 68567     
Total input vision tokens:               80256     
Total generated tokens:                  129952    
Total generated tokens (retokenized):    81784     
Request throughput (req/s):              1.14      
Input token throughput (tok/s):          1330.08   
Output token throughput (tok/s):         1161.42   
Total token throughput (tok/s):          2491.51   
Concurrency:                             48.33     
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   42245.86  
Median E2E Latency (ms):                 42871.80  
---------------Time to First Token----------------
Mean TTFT (ms):                          3458.57   
Median TTFT (ms):                        1652.74   
P99 TTFT (ms):                           8223.01   
--------------Time per Output Token---------------
Mean TPOT (ms):                          40.00     
Median TPOT (ms):                        40.37     
P99 TPOT (ms):                           70.99     
---------------Inter-Token Latency----------------
Mean ITL (ms):                           48.53     
Median ITL (ms):                         35.46     
P95 ITL (ms):                            106.41    
P99 ITL (ms):                            182.92    
Max ITL (ms):                            5496.99   
==================================================

[valarLip]

let me know once it's ready for review

[xytpai]

@valarLip Ready for review

[valarLip]

LGTM, nice job