[mrope][Qwen2-VL] Fix edge case where getting index of image/video token can potentially throw in default vl mrope implementation.

[huachenheli]

Purpose

In #23838, Gemini caught an edge case where checking image/video token in input_tokens does not guarantee input_tokens.index(image_token_id, st) won't throw, hence the missing try-catch block is needed. After discussion with @DarkLight1337 we want to add this check in the default vl mrope code as well.

Test Plan

Note that there's no meaningful perf impact since checking an item in the list is generally very fast:

(vllm) [[email protected] ~/github/vllm (mrope)]$ python -m timeit "99999 in range(100000)"
5000000 loops, best of 5: 84.3 nsec per loop
(vllm) [[email protected] ~/github/vllm (mrope)]$ python -m timeit "99999 in range(100000)"
5000000 loops, best of 5: 83 nsec per loop
(vllm) [[email protected] ~/github/vllm (mrope)]$ python -m timeit "99999 in range(100000)"
5000000 loops, best of 5: 84.3 nsec per loop

e2e benchmarks (reference only):
Without this PR:

============ Serving Benchmark Result ============
Successful requests:                     10        
Benchmark duration (s):                  3.75      
Total input tokens:                      156       
Total generated tokens:                  1667      
Request throughput (req/s):              2.67      
Output token throughput (tok/s):         444.86    
Total Token throughput (tok/s):          486.49    
---------------Time to First Token----------------
Mean TTFT (ms):                          943.28    
Median TTFT (ms):                        881.49    
P99 TTFT (ms):                           1103.42   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          11.93     
Median TPOT (ms):                        11.91     
P99 TPOT (ms):                           13.91     
---------------Inter-token Latency----------------
Mean ITL (ms):                           14.21     
Median ITL (ms):                         14.30     
P99 ITL (ms):                            17.47     
==================================================

With this PR:

============ Serving Benchmark Result ============
Successful requests:                     10        
Benchmark duration (s):                  3.69      
Total input tokens:                      156       
Total generated tokens:                  1642      
Request throughput (req/s):              2.71      
Output token throughput (tok/s):         445.57    
Total Token throughput (tok/s):          487.90    
---------------Time to First Token----------------
Mean TTFT (ms):                          888.17    
Median TTFT (ms):                        772.37    
P99 TTFT (ms):                           1079.04   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          12.35     
Median TPOT (ms):                        12.56     
P99 TPOT (ms):                           13.80     
---------------Inter-token Latency----------------
Mean ITL (ms):                           14.78     
Median ITL (ms):                         14.40     
P99 ITL (ms):                            16.96     
==================================================

Test Result

---

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.

[DarkLight1337]

Though this doesn't improve performance significantly, the code still becomes more robust. Thanks for working on this!