[FEAT] Support Triton MRoPE kernel

[tjtanaavllm]

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.

Purpose

Support Triton MRoPE kernel

Test Plan

ChartQA lmeval

Test Result

Baseline

{
    "explicit_prompt_relaxed_correctness": 0.8676,
    "anywhere_in_answer_relaxed_correctness": 0.8676
}

After

================================================================================
Metrics:
{
    "explicit_prompt_relaxed_correctness": 0.8648,
    "anywhere_in_answer_relaxed_correctness": 0.8648
}
================================================================================

MRoPE Torch

============ Serving Benchmark Result ============
Successful requests:                     1000      
Maximum request concurrency:             64        
Benchmark duration (s):                  572.61    
Total input tokens:                      94327    
Total generated tokens:                  114561    
Request throughput (req/s):              1.75      
Output token throughput (tok/s):         200.07    
Total Token throughput (tok/s):          364.80    
---------------Time to First Token----------------
Mean TTFT (ms):                          7619.61  
Median TTFT (ms):                        3190.75  
P99 TTFT (ms):                           32151.54  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          250.24    
Median TPOT (ms):                        278.42    
P99 TPOT (ms):                           379.93    
---------------Inter-token Latency----------------
Mean ITL (ms):                           252.85    
Median ITL (ms):                         17.00    
P99 ITL (ms):                            3913.67  
==================================================

MRope Triton

============ Serving Benchmark Result ============
Successful requests:                     1000      
Maximum request concurrency:             64        
Benchmark duration (s):                  547.88    
Total input tokens:                      94327    
Total generated tokens:                  114521    
Request throughput (req/s):              1.83      
Output token throughput (tok/s):         209.02    
Total Token throughput (tok/s):          381.19    
---------------Time to First Token----------------
Mean TTFT (ms):                          7377.66  
Median TTFT (ms):                        3118.68  
P99 TTFT (ms):                           31774.14  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          239.65    
Median TPOT (ms):                        269.29    
P99 TPOT (ms):                           348.70    
---------------Inter-token Latency----------------
Mean ITL (ms):                           241.30    
Median ITL (ms):                         16.84    
P99 ITL (ms):                            3343.22  
==================================================

Mrope Torch trace

MRope Triton Trace

(Optional) Documentation Update

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀