[qwen] Propagate onnxruntime-genai#2186: Qwen3.5-MoE text-only model type and TRT-RTX swiglu_limit

[Copilot]

Ports the builder-relevant changes from microsoft/onnxruntime-genai#2186 into the Qwen3.5-MoE builder. The upstream PR's other changes target C++ runtime files (model_type.h, config.cpp) and Python examples that are not part of this builder-only repository.

Changes

• Text-only model type (Qwen35MoeTextModel.__init__): select Qwen3_5_Moe_textForCausalLM when is_text_only, otherwise Qwen3_5_MoeForConditionalGeneration. The base builder strips the For… suffix and lowercases, so text-only LLM exports now emit genai-config model.type = qwen3_5_moe_text.
• TRT-RTX swiglu_limit fallback: when unset and ep == "trt-rtx", default to +inf to preserve "no clamp" behavior, since TRT-RTX QMoE builds require the attribute to be present.
• Test: added test_qwen3_5_moe_fp32_cpu_text_only_model_type asserting model.type == qwen3_5_moe_text for an exclude_embeds=False build; _build_model now forwards **extra_options.

self.model_type = "Qwen3_5_Moe_textForCausalLM" if self.is_text_only else "Qwen3_5_MoeForConditionalGeneration"

if self.moe_attrs.get("swiglu_limit") is None and self.ep == "trt-rtx":
    self.moe_attrs["swiglu_limit"] = float("inf")

The default (non-text-only) path is unchanged and still yields qwen3_5_moe.

CI fix: Qwen2.5-Omni 2D position_ids

The Fast Tests workflow was red on transformers ≥ 5 due to a pre-existing failure in the Qwen2.5-Omni thinker (RuntimeError: Invalid rank for input: position_ids Got: 2 Expected: 3). The thinker is exported as a standalone text decoder and driven by ORT-GenAI, which feeds 2D [B, S] position_ids, but the graph (inherited from Qwen25VLTextModel) declared 3D [3, B, S].

• Added an expand_position_ids flag on Qwen25VLTextModel (default False). When set, make_inputs_and_outputs declares a 2D position_ids input and make_mrope_position_ids expands it to 3D inside the graph (Unsqueeze + Tile, built once and reused per layer) for the mRoPE subgraph. Qwen25OmniThinkerModel sets the flag to True; genuine Qwen2.5-VL / Qwen3-VL keep 3D position_ids unchanged.
• Threaded an onnx_position_ids_2d parameter through run_mrope_vl_prefill_and_decode_check / run_vl_random_weights_test in ext_test_case.py (feeds 2D position_ids to ONNX while keeping the 3D feed to PyTorch), and updated the Qwen2.5-Omni tests to feed 2D position_ids.

[github-actions]

📂 Impacted Files

4 file(s) changed in this pull request:

🔵 modelbuilder/builders/qwen.py (+54 / -4)
🔵 modelbuilder/ext_test_case.py (+16 / -4)
🔵 tests/fast/test_random_qwen2_5_omni.py (+5 / -2)
🔵 tests/fast/test_random_qwen3_5_moe.py (+27 / -1)

[codecov]

Codecov Report

❌ Patch coverage is 92.59259% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.88%. Comparing base (f1a03b5) to head (52b4402).

Files with missing lines	Patch %	Lines
modelbuilder/builders/qwen.py	91.30%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #382      +/-   ##
==========================================
+ Coverage   78.87%   78.88%   +0.01%     
==========================================
  Files          27       27              
  Lines        8089     8113      +24     
==========================================
+ Hits         6380     6400      +20     
- Misses       1709     1713       +4     

Flag	Coverage Δ
fast-tests	78.88% <92.59%> (+0.02%)	⬆️
fast-tests-ort-nightly	78.83% <92.59%> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[xadupre]

@copilot fix build