[Model] Support DP for ViT on MiniCPM-V-4

[david6666666]

Purpose

Add option to run MiniCPM-V-4 vision encoder in data parallel manner while the main model is in TP. Can be enabled by flag: mm_encoder_tp_mode="data"

Test Plan & Result

lm_eval:

TP8(baseline):

lm_eval \
  --model vllm-vlm \
  --model_args "pretrained=openbmb/MiniCPM-V-4,tensor_parallel_size=8,gpu_memory_utilization=0.9,trust_remote_code=True" \
  --tasks chartqa \
  --limit 100 \
  --batch_size auto

| Tasks |Version|Filter|n-shot|     Metric      |   |Value |   |Stderr|
|-------|------:|------|-----:|-----------------|---|-----:|---|-----:|
|chartqa|      0|none  |     0|anywhere_accuracy|↑  |0.5640|±  |0.0099|
|       |       |none  |     0|exact_match      |↑  |0.4792|±  |0.0100|
|       |       |none  |     0|relaxed_accuracy |↑  |0.5640|±  |0.0099|

DP8:

lm_eval \
  --model vllm-vlm \
  --model_args "pretrained=openbmb/MiniCPM-V-4,mm_encoder_tp_mode=data,tensor_parallel_size=8,gpu_memory_utilization=0.9,trust_remote_code=True" \
  --tasks chartqa \
  --limit 100 \
  --batch_size auto

| Tasks |Version|Filter|n-shot|     Metric      |   |Value |   |Stderr|
|-------|------:|------|-----:|-----------------|---|-----:|---|-----:|
|chartqa|      0|none  |     0|anywhere_accuracy|↑  |0.5604|±  |0.0099|
|       |       |none  |     0|exact_match      |↑  |0.4740|±  |0.0100|
|       |       |none  |     0|relaxed_accuracy |↑  |0.5684|±  |0.0099|

Note: The accuracy is lower than the official standard, perhaps because chat_template was not applied, as MiniCPM did not provide it in the uploaded HF model.

performance:

TP8(baseline):

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

vllm serve openbmb/MiniCPM-V-4 \
    --gpu-memory-utilization 0.9 \
    --trust-remote-code \
    --tensor-parallel-size 8 \
    --host 0.0.0.0 \
    --port 20001

DP8:

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

vllm serve openbmb/MiniCPM-V-4 \
    --gpu-memory-utilization 0.9 \
    --mm_encoder_tp_mode data \
    --trust-remote-code \
    --tensor-parallel-size 8 \
    --host 0.0.0.0 \
    --port 20001

run benchmark_serving.py:

python3 benchmarks/benchmark_serving.py  \
    --backend openai-chat   \
    --model openbmb/MiniCPM-V-4 \
    --endpoint /v1/chat/completions   \
    --trust_remote_code \
    --dataset-name hf   \
    --dataset-path lmarena-ai/VisionArena-Chat   \
    --hf-split train   \
    --num-prompts 100 \
    --host localhost \
    --port 20001

result:

TP8(baseline):
============ Serving Benchmark Result ============
Successful requests:                     100       
Benchmark duration (s):                  26.45     
Total input tokens:                      6975      
Total generated tokens:                  9078      
Request throughput (req/s):              3.78      
Output token throughput (tok/s):         343.22    
Total Token throughput (tok/s):          606.93    
---------------Time to First Token----------------
Mean TTFT (ms):                          8593.27   
Median TTFT (ms):                        9483.75   
P99 TTFT (ms):                           21059.12  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          438.66    
Median TPOT (ms):                        170.73    
P99 TPOT (ms):                           4539.28   
---------------Inter-token Latency----------------
Mean ITL (ms):                           181.50    
Median ITL (ms):                         36.19     
P99 ITL (ms):                            3375.19   
==================================================

DP8:
============ Serving Benchmark Result ============
Successful requests:                     100       
Benchmark duration (s):                  17.58     
Total input tokens:                      6975      
Total generated tokens:                  8920      
Request throughput (req/s):              5.69      
Output token throughput (tok/s):         507.49    
Total Token throughput (tok/s):          904.32    
---------------Time to First Token----------------
Mean TTFT (ms):                          6977.63   
Median TTFT (ms):                        6616.90   
P99 TTFT (ms):                           13008.27  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          180.00    
Median TPOT (ms):                        107.37    
P99 TPOT (ms):                           832.16    
---------------Inter-token Latency----------------
Mean ITL (ms):                           101.05    
Median ITL (ms):                         35.87     
P99 ITL (ms):                            946.84    
==================================================

(Optional) Documentation Update

---

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.

[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.

🚀

[DarkLight1337]

Thanks, can you update the doc page with this model?

[DarkLight1337]

Thanks, LGTM

[DarkLight1337]

cc @jio-H you may consider using this feature

[ZJY0516]

Sorry to bother you, but I was wondering why it's 3 * self.q_size instead of self.head_dim?

[CarrotShoo]

Perhaps you misunderstood the meaning of the parameters? In ReplicatedRinear, 3 * self.q_size is the output_size, self.head_dim in QKVParallelLinear means head_size. Hope my answer has been helpful to you.

[ZJY0516]

Thanks for explaining! Got it now.