Validate Qwen vision embedding output

[Akshay Sonawane (apsonawane)]

This pull request improves the robustness of vision embedding injection in the Qwen2.5-VL model by centralizing and expanding validation of tensor shapes and element types, and by adding targeted unit tests. The main changes ensure that runtime errors due to shape mismatches or invalid data types are caught early and handled with clear error messages.

Validation improvements:

• Added a new function ValidateVisionEmbeddingShapes in qwen_vl_model.h to check that embeddings and vision feature tensors have the correct rank, matching dimensions, and sufficient size for all input tokens.
• Updated InjectVisionEmbeddings in qwen_vl_model.cpp to call ValidateVisionEmbeddingShapes and to explicitly check that the embeddings tensor contains float elements, throwing descriptive exceptions otherwise. [1] [2]

Testing:

• Added a new unit test QwenVisionEmbeddingShapeValidation in model_tests.cpp to verify that valid shapes pass and invalid shapes throw the expected runtime errors.

[Copilot]

Pull request overview

This PR strengthens Qwen2.5-VL vision embedding injection by centralizing validation of embedding/vision tensor shapes and adding a focused unit test to ensure invalid shapes fail fast with clear runtime errors.

Changes:

• Added Generators::ValidateVisionEmbeddingShapes to validate embedding/vision tensor ranks, dims, and token capacity.
• Updated Qwen2_5_VL_PipelineState::InjectVisionEmbeddings to reuse the centralized shape validation and to reject non-float embedding outputs early.
• Added a C++ unit test covering valid/invalid vision embedding shape scenarios.

Reviewed changes

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

File	Description
test/cpp/model_tests.cpp	Adds a unit test for the new shape validation helper.
src/models/qwen_vl_model.h	Introduces centralized vision embedding shape validation helper.
src/models/qwen_vl_model.cpp	Uses centralized validation and adds explicit float element-type check before embedding injection.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.