使用vllm部署gptqmodel或auto-aptq量化的微调后的qwen3-vl时报错：KeyError: 'layers.0.mlp.down_proj.g_idx'

Describe the bug
What the bug is, and how to reproduce, better with screenshots(描述bug以及复现过程，最好有截图)

Your hardware and system info
Write your system info like CUDA version/system/GPU/torch version here(在这里给出硬件信息和系统信息，如CUDA版本，系统，GPU型号和torch版本等)

Additional context
Add any other context about the problem here(在这里补充其他信息)

ms-swift main分支最新
vllm 0.11.0

量化使用的ms-swift中的gptq以及qptq-v2

部署指令：
CUDA_VISIBLE_DEVICES=0,1,2,3 python3.10 -m vllm.entrypoints.openai.api_server --served-model-name pretrain_8b --model xxx-Instruct-DPO-INT8-GPTQ-V2 --gpu-memory-utilization 0.8 --tensor-parallel-size 4 --port 8799 --enforce-eager

Start by reproducing the reported command through vllm.entrypoints.openai.api_server with vllm 0.11.0 and the ms-swift main branch, using the GPTQ or QPTQ-v2 model. Trace the loading failure for layers.0.mlp.down_proj.g_idx and compare the quantized model metadata with what the Qwen3-VL loader expects; done means the model loads successfully or the incompatibility is clearly documented.