Why Qwen2VL can generate normally with zero visual_features

[YSF-A]

With the default image, Qwen2VL can generate normally.

When I replace visual_features with zero tensor (add visual_features[:,:]=0 after https://github.com/NVIDIA/TensorRT-LLM/blob/main/tensorrt_llm/runtime/multimodal_model_runner.py#L699) , it can also generate normally and the result is the same.

Why Qwen2VL can generate normally with zero visual_features? Is the information of image included by somewhere else?

[sunnyqgg]

Hi @YSF-A ,
Could you try it with the latest main and if it still exists, please let me know it