Enable Qwen3.5 35B-A3B MoE text export for TRT-RTX EP

[yen-shi]

Summary

Enable Qwen3.5 35B-A3B MoE text-only export and inference with the TRT-RTX execution provider.

• Emit a decoder-only GenAI model type for TRT-RTX text-only Qwen3.5-MoE exports.
• Keep the existing qwen3_5_moe VLM path intact for non-TRT-RTX/non-text behavior.
• Preserve model-provided swiglu_limit when available, and fall back to +inf when the config omits it.
• Allow Python examples to use the GenAI provider name NvTensorRtRtx while registering the external ORT EP library as NvTensorRTRTXExecutionProvider.

Validation

• Exported Qwen3.5 35B-A3B MoE INT4 QDQ text-only model with TRT-RTX EP.
• Verified generated config:
  • model.type = qwen3_5_moe_text
  • decoder file is text.onnx
  • TRT-RTX provider options include enable_cuda_graph = 1
• Verified exported ONNX:
  • QMoE node count: 40
  • no QMoE nodes missing swiglu_limit
• Ran TRT-RTX EP inference with model-qa.py using -e NvTensorRtRtx.
• Smoke-tested the same -e NvTensorRtRtx alias on Qwen3.5 0.8B and 9B TRT-RTX packages.

Notes

• This change does not remove or alter the existing qwen3_5_moe VLM model registration.
• The TRT-RTX text-only path is selected only for the TRT-RTX export path with text-only settings.
• The Python example alias is intentionally minimal: it only maps the GenAI provider name to the external ORT EP library name during registration.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Updates Qwen3.5-MoE builder logic and EP registration to better support TRT-RTX text-only execution, and adds a new model type identifier for LLM classification.

Changes:

• Select a different Qwen3.5-MoE model_type when running TRT-RTX with text-only.
• Populate swiglu_limit in MoE attributes with a fallback lookup/default.
• Add qwen3_5_moe_text to the C++ LLM model type allowlist and accept NvTensorRtRtx as an EP alias in the Python examples.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
src/python/py/models/builders/qwen.py	Adjusts Qwen3.5-MoE model type selection for TRT-RTX text-only and adds swiglu_limit handling.
src/models/model_type.h	Adds qwen3_5_moe_text to the LLM classification list.
examples/python/common.py	Adds NvTensorRtRtx EP option and maps it to the provider library name when registering.

[anskumar01]

@kunal-vaishnavi , @baijumeswani , can you please review

[yen-shi]

Hi @kunal-vaishnavi,
Do you know how to pass or restart the pipeline?
One windows pipeline failed, but I don't see clear error message from the logs.