Bugging Results? Repetitive Outputs and Long Inference Times with LLaVA-Video-7B-Qwen2

[Rachel0901]

I’ve been using the pre-trained LLaVA-Video-7B-Qwen2 model for some inference tasks and have noticed a couple of issues that I’d like to share to see if others have experienced something similar:

• Repetitive Outputs: Some outputs contain repetitive content, repeating the same response up to 50 times.
• Extended Inference Times: For outputs with repetitive content, the inference times are noticeably longer than others.

Has anyone else noticed similar issues with repetitive outputs using this model? Additionally, do you have any suggestions for mitigating these problems (e.g., changing the instruction prompt, adjusting decoding parameters, or modifying the model configuration)?