[Bug] Qwen3-VL-30B-A3B NVFP4 quantization fails: hidden_size AttributeError in QuantQwen3VLMoeTextExperts

Environment

• nvidia-modelopt: 0.44.0 (latest PyPI + tested with latest source from main)
• TensorRT-Edge-LLM: 0.8.0
• Platform: NVIDIA Jetson AGX Thor (SM110, JetPack 7.2, CUDA 13.2)
• Model: Qwen/Qwen3-VL-30B-A3B-Instruct (HuggingFace)
• Python: 3.12

Problem

NVFP4 quantization of Qwen3-VL-30B-A3B-Instruct fails with a shape mismatch
in QuantQwen3VLMoeTextExperts during the MoE expert forward pass.

Error chain

1. AttributeError: hidden_size in QuantQwen3VLMoeTextExperts
2. mat1/mat2 shape mismatch in MoE expert forward

The error originates in modelopt's handling of this model's fused
gate_up_proj / refactored expert dimensions which do not match the
expected shapes in the MoE quantization code.

Steps to reproduce

tensorrt-edgellm-quantize-llm \
  --model_dir ~/models/Qwen3-VL-30B-A3B-Instruct/ \
  --output_dir ~/models/qwen3-vl-30b-nvfp4/onnx/ \
  --quantization nvfp4

Start by reproducing the provided tensorrt-edgellm-quantize-llm command with Qwen/Qwen3-VL-30B-A3B-Instruct and NVFP4. Inspect QuantQwen3VLMoeTextExperts and the modelopt handling of fused gate_up_proj and refactored expert dimensions. Done means quantization completes without the hidden_size AttributeError or MoE expert shape mismatch.