Remove mrope position sync

[timmy-feng]

Motivation

Running _compute_mrope_positions() on the forward batch during decode incurs an unnecessary host sync. This prevents us from achieving full overlap on Qwen2.5 VL series among other VLMs.

Modifications

We avoid moving mrope_position_delta to the host by computing position indices directly on the GPU.

Accuracy Tests

Before

Benchmark time: 76.03708224199909
answers saved to: ./answer_sglang.json
Evaluating...
{'Accounting': {'acc': 0.433, 'num': 30},
 'Agriculture': {'acc': 0.533, 'num': 30},
 'Architecture_and_Engineering': {'acc': 0.367, 'num': 30},
 'Art': {'acc': 0.567, 'num': 30},
 'Art_Theory': {'acc': 0.7, 'num': 30},
 'Basic_Medical_Science': {'acc': 0.667, 'num': 30},
 'Biology': {'acc': 0.367, 'num': 30},
 'Chemistry': {'acc': 0.267, 'num': 30},
 'Clinical_Medicine': {'acc': 0.4, 'num': 30},
 'Computer_Science': {'acc': 0.3, 'num': 30},
 'Design': {'acc': 0.633, 'num': 30},
 'Diagnostics_and_Laboratory_Medicine': {'acc': 0.433, 'num': 30},
 'Economics': {'acc': 0.3, 'num': 30},
 'Electronics': {'acc': 0.133, 'num': 30},
 'Energy_and_Power': {'acc': 0.4, 'num': 30},
 'Finance': {'acc': 0.267, 'num': 30},
 'Geography': {'acc': 0.3, 'num': 30},
 'History': {'acc': 0.733, 'num': 30},
 'Literature': {'acc': 0.833, 'num': 30},
 'Manage': {'acc': 0.3, 'num': 30},
 'Marketing': {'acc': 0.5, 'num': 30},
 'Materials': {'acc': 0.433, 'num': 30},
 'Math': {'acc': 0.333, 'num': 30},
 'Mechanical_Engineering': {'acc': 0.367, 'num': 30},
 'Music': {'acc': 0.233, 'num': 30},
 'Overall': {'acc': 0.452, 'num': 900},
 'Overall-Art and Design': {'acc': 0.533, 'num': 120},
 'Overall-Business': {'acc': 0.36, 'num': 150},
 'Overall-Health and Medicine': {'acc': 0.533, 'num': 150},
 'Overall-Humanities and Social Science': {'acc': 0.7, 'num': 120},
 'Overall-Science': {'acc': 0.327, 'num': 150},
 'Overall-Tech and Engineering': {'acc': 0.362, 'num': 210},
 'Pharmacy': {'acc': 0.6, 'num': 30},
 'Physics': {'acc': 0.367, 'num': 30},
 'Psychology': {'acc': 0.6, 'num': 30},
 'Public_Health': {'acc': 0.567, 'num': 30},
 'Sociology': {'acc': 0.633, 'num': 30}}
eval out saved to ./val_sglang.json
Overall accuracy: 0.452

After

Benchmark time: 77.83271087400044
answers saved to: ./answer_sglang.json
Evaluating...
{'Accounting': {'acc': 0.433, 'num': 30},
 'Agriculture': {'acc': 0.533, 'num': 30},
 'Architecture_and_Engineering': {'acc': 0.367, 'num': 30},
 'Art': {'acc': 0.567, 'num': 30},
 'Art_Theory': {'acc': 0.7, 'num': 30},
 'Basic_Medical_Science': {'acc': 0.667, 'num': 30},
 'Biology': {'acc': 0.367, 'num': 30},
 'Chemistry': {'acc': 0.267, 'num': 30},
 'Clinical_Medicine': {'acc': 0.4, 'num': 30},
 'Computer_Science': {'acc': 0.3, 'num': 30},
 'Design': {'acc': 0.6, 'num': 30},
 'Diagnostics_and_Laboratory_Medicine': {'acc': 0.433, 'num': 30},
 'Economics': {'acc': 0.3, 'num': 30},
 'Electronics': {'acc': 0.133, 'num': 30},
 'Energy_and_Power': {'acc': 0.4, 'num': 30},
 'Finance': {'acc': 0.267, 'num': 30},
 'Geography': {'acc': 0.333, 'num': 30},
 'History': {'acc': 0.733, 'num': 30},
 'Literature': {'acc': 0.833, 'num': 30},
 'Manage': {'acc': 0.3, 'num': 30},
 'Marketing': {'acc': 0.5, 'num': 30},
 'Materials': {'acc': 0.433, 'num': 30},
 'Math': {'acc': 0.367, 'num': 30},
 'Mechanical_Engineering': {'acc': 0.367, 'num': 30},
 'Music': {'acc': 0.233, 'num': 30},
 'Overall': {'acc': 0.453, 'num': 900},
 'Overall-Art and Design': {'acc': 0.525, 'num': 120},
 'Overall-Business': {'acc': 0.36, 'num': 150},
 'Overall-Health and Medicine': {'acc': 0.533, 'num': 150},
 'Overall-Humanities and Social Science': {'acc': 0.7, 'num': 120},
 'Overall-Science': {'acc': 0.34, 'num': 150},
 'Overall-Tech and Engineering': {'acc': 0.362, 'num': 210},
 'Pharmacy': {'acc': 0.6, 'num': 30},
 'Physics': {'acc': 0.367, 'num': 30},
 'Psychology': {'acc': 0.6, 'num': 30},
 'Public_Health': {'acc': 0.567, 'num': 30},
 'Sociology': {'acc': 0.633, 'num': 30}}
eval out saved to ./val_sglang.json
Overall accuracy: 0.453

Benchmarking and Profiling

Before

============ Serving Benchmark Result ============
Backend:                                 sglang    
Traffic request rate:                    inf       
Max request concurrency:                 16        
Successful requests:                     200       
Benchmark duration (s):                  79.12     
Total input tokens:                      13654     
Total generated tokens:                  204800    
Total generated tokens (retokenized):    124526    
Request throughput (req/s):              2.53      
Input token throughput (tok/s):          172.58    
Output token throughput (tok/s):         2588.58   
Total token throughput (tok/s):          2761.16   
Concurrency:                             15.40     
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   6091.23   
Median E2E Latency (ms):                 6082.71   
---------------Time to First Token----------------
Mean TTFT (ms):                          97.26     
Median TTFT (ms):                        65.02     
P99 TTFT (ms):                           511.28    
---------------Inter-Token Latency----------------
Mean ITL (ms):                           5.86      
Median ITL (ms):                         5.89      
P95 ITL (ms):                            6.10      
P99 ITL (ms):                            6.34      
Max ITL (ms):                            38.79     
==================================================

After

============ Serving Benchmark Result ============
Backend:                                 sglang    
Traffic request rate:                    inf       
Max request concurrency:                 16        
Successful requests:                     200       
Benchmark duration (s):                  65.90     
Total input tokens:                      13654     
Total generated tokens:                  204800    
Total generated tokens (retokenized):    126196    
Request throughput (req/s):              3.03      
Input token throughput (tok/s):          207.19    
Output token throughput (tok/s):         3107.72   
Total token throughput (tok/s):          3314.91   
Concurrency:                             15.39     
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   5071.40   
Median E2E Latency (ms):                 5077.57   
---------------Time to First Token----------------
Mean TTFT (ms):                          59.60     
Median TTFT (ms):                        61.08     
P99 TTFT (ms):                           73.16     
---------------Inter-Token Latency----------------
Mean ITL (ms):                           4.90      
Median ITL (ms):                         4.92      
P95 ITL (ms):                            5.24      
P99 ITL (ms):                            5.41      
Max ITL (ms):                            31.61     
==================================================

Repro Script

The following script was run on 1xH200.

#! /bin/bash

# Start SGLang server
python -m sglang.launch_server \
    --model-path Qwen/Qwen2.5-VL-3B-Instruct \
    --port 30000 &
PID=$!

# Wait for server to start
while ! curl -s http://localhost:30000/health > /dev/null; do
    sleep 1
done

# Run accuracy benchmark
python benchmark/mmmu/bench_sglang.py \
    --port 30000 \
    --concurrency 16

# Flush cache
curl -X POST http://localhost:30000/flush_cache

# Run latency benchmark
python -m sglang.bench_serving \
    --backend sglang \
    --dataset-name mmmu \
    --num-prompts 200 \
    --max-concurrency 16

# Kill server
kill $PID

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.

[gemini-code-assist]

Summary of Changes

Hello @timmy-feng, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request aims to enhance the performance of multimodal models, particularly during the decode phase, by optimizing the computation of rotary embedding positions. It eliminates an unnecessary host-device synchronization by ensuring that position indices are calculated directly on the GPU, thereby improving computational overlap and overall efficiency.

Highlights

• Performance Improvement: Addresses a performance bottleneck caused by host synchronization during mrope_position computation in the decode phase.
• GPU-centric Computation: Refactors the logic to compute mrope_position_deltas and subsequent position indices entirely on the GPU, avoiding costly data transfers to the host.
• VLM Optimization: Specifically targets and benefits Visual Language Models (VLMs) like the Qwen2.5 VL series by enabling better computational overlap.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request refactors the _compute_mrope_positions method to avoid unnecessary host synchronization during the decode phase. The changes correctly move the position index computation directly to the GPU, which should improve performance as intended. The new implementation is also more concise and readable. The logic appears sound and is a good optimization.