Fix Quark/AWQ weight loading for Qwen3-VL-4B text model

[anilmartha]

Summary

Fixes Quark/AWQ weight loading for the Qwen3-VL-4B text model when building the ONNX model from a quantized Hugging Face checkpoint.
Previously:

• Qwen3VLTextModel.load_weights always went through Qwen3VLForConditionalGeneration.from_pretrained, which doesn't understand quantized checkpoints (Quark / AWQ / GPTQ) or .gguf files.
• The shared QuantizedModel loader didn't handle VLM checkpoint conventions (vision-tower weights, model.language_model.* prefix) or Quark's weight_quantizer.{scale,zero_point} naming.

Changes

src/python/py/models/builders/qwen.py

• In Qwen3VLTextModel.load_weights, route quantized models (self.quant_type is not None) and .gguf inputs to the base class implementation, which loads weights via QuantizedModel. Plain HF FP checkpoints continue to use Qwen3VLForConditionalGeneration.from_pretrained as before.
  src/python/py/models/quantized_model.py
• Skip vision-tower tensors (model.visual.*, model.vision.*, visual.*) so the LLM weight loader only consumes language-model weights.
• Normalize the VLM prefix model.language_model.* to model.* so the existing LLM regex/tensor-parsing logic keeps working unchanged.
• Normalize Quark naming:
  • *.weight_quantizer.scale → *.weight_scale
  • *.weight_quantizer.zero_point → *.weight_zero_point

[anilmartha]

@microsoft-github-policy-service agree company="AMD"

[Copilot]

Pull request overview

Fixes Qwen3-VL-4B text model weight loading when building ONNX from quantized Hugging Face checkpoints (Quark/AWQ/GPTQ) or GGUF inputs by routing those cases through the shared quantized/GGUF loader and expanding the quantized tensor name-normalization logic to handle VLM checkpoint conventions.

Changes:

• Route Qwen3VLTextModel.load_weights to the base-class loader for quantized checkpoints and .gguf inputs (keeping the existing HF FP loading path unchanged).
• Extend QuantizedModel weight-key handling to (a) skip vision-tower tensors in VLM checkpoints, (b) normalize model.language_model.* to model.*, and (c) normalize Quark weight_quantizer.{scale,zero_point} naming to the existing weight_{scale,zero_point} convention.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/python/py/models/quantized_model.py	Adds VLM/Quark-specific tensor-key filtering and normalization so existing LLM parsing logic can consume Qwen3-VL language weights.
src/python/py/models/builders/qwen.py	Ensures Qwen3-VL text builder uses the shared quantized/GGUF loading path for non-FP checkpoints.