input prompts only, test qwen3 omni.

[xipingyan]

This pull request adds a new configuration YAML file for the Qwen3-Omni model pipeline in the samples/cpp/module_genai/config_yaml directory. The configuration defines the structure and flow for prompt encoding, LLM inference, and result formatting, enabling streamlined integration and testing of the Qwen3-Omni model.

New Qwen3-Omni pipeline configuration:

• Added config_prompt.yaml to specify the pipeline modules and their parameters for the Qwen3-Omni model, including prompt encoding, LLM inference, and result collection.
• Defined module chaining and device allocation (GPU) for both the TextEncoderModule and LLMInferenceSDPAModule, supporting efficient model execution.
• Provided explicit model paths and parameters for both encoder and LLM inference modules, facilitating reproducible testing and deployment.
• Established output structure via the ResultModule, ensuring generated text is properly collected and formatted.

ISSUE:

[INFO] Running module: pipeline_params
[INFO]     Pass prompt to output port
[INFO] Running module: prompt_encoder
[INFO] Running module: llm
[ERROR] Exception from src/inference/src/cpp/infer_request.cpp:224:
Check 'TRShape::broadcast_merge_into(result_shape, input_shapes[input_port], broadcast_spec)' failed at src/core/shape_inference/include/select_shape_inference.hpp:40:
While validating node 'opset1::Select Select_401311 () -> ()' with friendly_name 'Select_401311':
Argument shapes are inconsistent.

[Copilot]

Pull request overview

This PR adds a new text-only (prompt-only) pipeline configuration YAML for the Qwen3-Omni model. The configuration strips away the vision/audio modules from the full config.yaml to create a simplified pipeline that only processes text prompts through the TextEncoderModule → LLMInferenceSDPAModule → ResultModule chain, similar to the existing Qwen3.5-0.8B/config_text.yaml.

Changes:

• Added config_prompt.yaml for Qwen3-Omni with a text-only pipeline (prompt encoding → LLM inference → result collection), omitting vision and audio preprocessing stages.
• Defined GPU device allocation for both the TextEncoderModule and LLMInferenceSDPAModule with model paths pointing to the Qwen3-Omni-4B test models.
• Configured the pipeline to generate up to 512 tokens and collect the output through a ResultModule.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The prompt_encoder is missing the mask output declaration. The TextEncoderModule::run() method unconditionally writes to this->outputs["mask"].data for the QWEN3_OMNI model type (see src/cpp/src/module_genai/modules/md_text_encoder.cpp:159). While this won't crash (the map auto-inserts), it's inconsistent with every other TextEncoderModule config in the codebase — for example, the analogous text-only config at samples/cpp/module_genai/config_yaml/Qwen3.5-0.8B/config_text.yaml:21-22 and the sibling config.yaml at samples/cpp/module_genai/config_yaml/Qwen3-Omni/config.yaml:48-49 both declare the mask output. Add the mask output to maintain consistency and properly document the module's outputs.

Suggested change

	type: "OVTensor"
	type: "OVTensor"
	- name: "mask"
	type: "OVTensor"

[Copilot]

A blank line is missing between the pipeline_params and prompt_encoder module definitions. All other pipeline YAML configs in this directory consistently use blank lines to separate top-level modules (e.g., Qwen3.5-0.8B/config_text.yaml:10, Qwen3-Omni/config.yaml:12). Add a blank line after line 9 for consistency.

Suggested change

	type: "String"
	type: "String"