FEAT: [model] Support Qwen3-VL

[Jun-Howie]

What's New

1. Model Support

   • Added support for Qwen3-VL Instruct/Thinking models.

2. Quantized Models

   • Provided additional FP8 and AWQ quantized models.
   • Fully tested and supported.

3. Tool Calling

   • Added support for tool_call, consistent with the Qwen3 series.

4. Frontend UI

   • Introduced the --enable-expert-parallel parameter.
   • Aligned with vllm/core.py, enabling expert parallelism so that each GPU is allocated the same number of experts.

---

Dependencies (vLLM Backend)

pip install git+https://github.com/huggingface/transformers
pip install qwen-vl-utils==0.0.14

# Recommended (if the stable version is not working, use nightly)
# pip install 'vllm>0.10.2' 
uv pip install -U vllm \
    --torch-backend=auto \
    --extra-index-url https://wheels.vllm.ai/nightly

Usage Examples

2 × A100-80G (AWQ)

vllm serve \
    ./Qwen3-VL-235B-A22B-Instruct-AWQ \
    --enable-expert-parallel \
    --max-model-len 32768 \
    --tensor-parallel-size 2

4 × A100-80G (FP8)

vllm serve \
   ./Qwen3-VL-235B-A22B-Instruct-FP8 \
   --enable-expert-parallel \
   --max-model-len 32768 \
   --tensor-parallel-size 4

Test Results

[qinxuye]

LGTM