Add text-only mode support for Qwen 3.5 model builder

[apsonawane]

Description

Adds support for running Qwen 3.5 as a standalone text-only LLM (without vision/embedding pipelines).

Changes

• 2D position_ids support: When exclude_embeds=false, the builder creates a 2D [B, S] position_ids graph input and internally expands it to 3D [3, B, S] for mRoPE compatibility. This allows the standard onnxruntime-genai runtime to provide position_ids without requiring the multimodal pipeline.
• Tokenizer regex fix: Added save_processing override that patches unsupported \p{M} (Unicode Mark category) from tokenizer regex patterns after export. The C++ std::regex engine in onnxruntime-extensions does not support this Unicode property class.
• genai_config correctness: The internal 3D expanded tensor name is stored in a separate _pos_ids_3d attribute so self.input_names["position_ids"] remains as "position_ids" — ensuring genai_config.json references the actual graph input.

Usage

python3 -m onnxruntime_genai.models.builder \
  --model_name Qwen/Qwen3.5-2B \
  --precision int4 \
  --execution_provider cuda \
  --output /path/to/output \
  --extra_options exclude_embeds=false prune_lm_head=true int4_algo_config=k_quant_last

[Copilot]

Pull request overview

Adds support for building and running Qwen 3.5 as a standalone text-only LLM (without multimodal embedding/vision pipeline), including runtime-side fixes to avoid incorrectly injecting input_ids into decoders that only accept inputs_embeds.

Changes:

• Add “text-only mode” to the Qwen 3.5 builder, including 2D position_ids support with internal expansion for mRoPE and a tokenizer-regex post-export patch.
• Fix multimodal runtime logic to check decoder inputs against decoder-only session metadata (avoids false positives from the embedding session).
• Register qwen3_5_text as an LLM model type in C++.

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	Adds text-only build path for Qwen 3.5, adjusts position_ids handling, and patches exported tokenizer regex.
src/models/multi_modal.cpp	Refines decoder input detection to avoid injecting input_ids when the decoder session doesn’t accept it.
src/models/model_type.h	Adds qwen3_5_text to the LLM model-type allowlist.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

[kunal-vaishnavi]

Why do the position ids have to differ in text-only mode vs. vision-language mode? It should be an identical decoder besides the input_ids vs inputs_embeds input.

[apsonawane]

The difference exists because of who provides them, not the decoder itself.

1. VL mode: The multimodal pipeline computes 3D [3, B, S] position IDs externally — image patches need distinct spatial (H/W) positions vs text tokens (temporal only).
2. Text-only mode: The onnxruntime-genai runtime only generates standard 2D [B, S] position IDs. For pure text, all 3 mRoPE dimensions are identical (sequential), so we Tile [B, S] → [3, B, S] in the graph.

To eliminate the Tile, we'd need to:

1. Add qwen3_5_text to IsQwenVLFamily() in model_type.h
2. Guard vision token config reads in Qwen2VLPositionInputs for the no-vision case in position_inputs.cpp
3. Remove the is_text_only branching in the Python builder

There is no perf regression, adding it in runtime will add code complexity I would say. Also, regarding above changes keeping text model out of VL model type makes sense to me.