[Kernel][Performance] Add Triton kernel for Qwen3-VL interleaved MRoPE

[Isotr0py]

Purpose

• Following PR for [Model] Support Qwen3-VL Model Series #24727, implement corresponding Triton kernel for interleaved MRoPE.

Test Plan

pytest -s -v tests/kernels/core/test_mrope.py

Test Result

Test should pass

---

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.

[Isotr0py]

Benchmark

server

vllm serve /home/mozf/LLM/Qwen3-VL-4B-Instruct/

client

vllm bench serve  \
--backend openai-chat   \
--endpoint-type openai-chat \
--model /home/mozf/LLM/Qwen3-VL-4B-Instruct/   \
--endpoint /v1/chat/completions   \
--dataset-name hf   \
--dataset-path lmarena-ai/VisionArena-Chat   \
--hf-split train   \
--num-prompts 200 \
--max-concurrency 64

Results

Main

============ Serving Benchmark Result ============
Successful requests:                     993       
Maximum request concurrency:             64        
Benchmark duration (s):                  186.79    
Total input tokens:                      94247     
Total generated tokens:                  119911    
Request throughput (req/s):              5.32      
Output token throughput (tok/s):         641.94    
Total Token throughput (tok/s):          1146.49   
---------------Time to First Token----------------
Mean TTFT (ms):                          1005.91   
Median TTFT (ms):                        708.54    
P99 TTFT (ms):                           5738.68   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          95.17     
Median TPOT (ms):                        90.26     
P99 TPOT (ms):                           290.06    
---------------Inter-token Latency----------------
Mean ITL (ms):                           91.23     
Median ITL (ms):                         29.79     
P99 ITL (ms):                            455.95    
==================================================

PR

============ Serving Benchmark Result ============
Successful requests:                     990       
Maximum request concurrency:             64        
Benchmark duration (s):                  185.86    
Total input tokens:                      93035     
Total generated tokens:                  119586    
Request throughput (req/s):              5.33      
Output token throughput (tok/s):         643.43    
Total Token throughput (tok/s):          1144.00   
---------------Time to First Token----------------
Mean TTFT (ms):                          959.61    
Median TTFT (ms):                        703.75    
P99 TTFT (ms):                           5533.73   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          94.51     
Median TPOT (ms):                        90.73     
P99 TPOT (ms):                           234.72    
---------------Inter-token Latency----------------
Mean ITL (ms):                           91.84     
Median ITL (ms):                         30.29     
P99 ITL (ms):                            458.38    
==================================================

[DarkLight1337]

Can you also check lm-eval?

[ywang96]

@Isotr0py @DarkLight1337 BTW let's not post the actual eval numbers since the model hasn't been released - just make sure the results match

[Isotr0py]

just make sure the results match

Oh, I just realized there is no generation tests for Qwen3-VL yet. Let me add ones later today (after my seminar)

[Isotr0py]

Have confirmed Qwen3-VL generation test added by #25185 can still pass with Triton kernel.