Meaningless response for video demo

[TRS07170]

I tried to run the video_demo.sh script on my own video. I only modified the video path with out changing any other parameters:

bash scripts/video/demo/video_demo.sh lmms-lab/LLaVA-NeXT-Video-7B-DPO vicuna_v1 32 2 average no_token True /path/to/my/video.mp4

Then the output becomes empty, like this:

If I reduce the number sampled frames to 10 or 20, it generate at least readable responses. But if I increase the number of sampled frames to 30 and run the following command:

bash scripts/video/demo/video_demo.sh lmms-lab/LLaVA-NeXT-Video-7B-DPO vicuna_v1 30 2 average no_token True /path/to/my/video.mp4

then the model starts to output meaningless responses:

I wonder how could this happen? Did set any parameters wrong?