videochat-7b returns garbled code

[ssfgunner]

videochat-7b returns garbled code

[Andy1621]

@ssfgunner Thanks for your problem! The wrong weights of vicuna-7b may cause your strange results.

As for the log, it's expected that those keys are not loaded by weight eva_vit_g.pth, since those keys belong to the extra GMHRA modules. Besides, those keys are loaded by the weight of our VideoChat.

To fix your bug, you can refer to the Vision-CAIR/MiniGPT-4#12 in MiniGPT-4. I have fixed it via correctly loading the weight of Vicuna. You may only load the delta weight.

[ssfgunner]

@Andy1621 Thanks for your reply! You are right, I accidentally only loaded delta weight, I solved this problem by loading the correct weight, thanks again!