Support qwen3-omni with DP Encoder

[apinge]

Motivation

Add support for Qwen3-Omni with vision model DP based on #13724

Modifications

Accuracy Tests

Tested Qwen/Qwen3-Omni-30B-A3B-Instruct

with dp

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
mmmu_val	0	none	0	mmmu_acc	↑	0.5867	±	N/A

without dp

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
mmmu_val	0	none	0	mmmu_acc	↑	0.59	±	N/A

Benchmarking and Profiling

On H20, we benchmarked the server with the --mm-enable-dp-encoderoption enabled against the baseline performance without any code modifications. We found that enabling --mm-enable-dp-encoder reduces TTFT for multi-modal workloads under some testing conditions.

SGLANG_VLM_CACHE_SIZE_MB=0  
python3 -m sglang.launch_server \
         --model-path /root/workspace/Qwen3-Omni-30B-A3B-Instruct/   \
         --host localhost \
         --port 40000 \
         --tp-size 4 \
        --trust-remote-code \
         --mem-fraction-static 0.85  \
        --disable-radix-cache   \
        --cuda-graph-max-bs 32 \
        --mm-attention-backend fa3 \
        --attention-backend flashinfer\
        --mm-enable-dp-encoder

tp	Img count	Img resolution	TTFT base(ms)	TTFT with dp(ms)	improment
4	4	960x1280	6733.54	6543.15	-2.83%
4	8	960x1280	16671.97	14916.71	-10.53%%

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.
• Follow the SGLang code style guidance.
• Work with maintainers to merge your PR. See the PR Merge Process