search

haotian-liu / LLaVA

[NeurIPS'23 Oral] Visual Instruction Tuning (LLaVA) built towards GPT-4V level capabilities and beyond.
https://llava.hliu.cc
Apache License 2.0
19.3k stars 2.12k forks source link

repetition penalty #836

Open SuXuping opened 10 months ago

SuXuping commented 10 months ago

Question

when i try to use repetition_penalty to avoid repeat answer, i met this error "cuda error:device-side assert triggered". After my debug, i found that the input_ids include -200 which is default image token. i guess this is the reason of this " cuda error:device-side assert triggered". can you fix this error?

Jeckinchen commented 6 months ago

Hello，have you solved this problem? I also encountered the problem of repeated responses after fine-tuning the model.

redscv commented 4 months ago

I had the same error after set repetition_penalty > 0

LanYu4qz commented 3 months ago

I have also encountered this issue. Who can solve it?

SuXuping commented 3 months ago

I have also encountered this issue. Who can solve it? change -200 to a new positive num

LanYu4qz commented 3 months ago

I have also encountered this issue. Who can solve it? change -200 to a new positive num

Could you give me some guidance on how to do it specifically?

zhangsha1024 commented 3 months ago

same here