AttackVLM
AttackVLM copied to clipboard
Segmentation fault when doing `clip_model.encode_image`
Thanks for the amazing work. I am now trying to run the _train_adv_img_trans.py code and get an unexpected segmentation error. After locating the error location, I believe it is from the clip_model.encode_image() part.
Here is the detail of the error.
can let this function instead forward_encoder_image in blip2?
It is usually due to the package version or your installation environment (e.g. cuda). see if this can help: https://github.com/ggerganov/llama.cpp/issues/4702