Hello, I am a graduate student and using video inference.
After inputting a video into the model, I found that in the 18th encoder layer of the MplugOwlVisionEncoderLayer, during mlp, after QuickGELU activation function, the values of hidden_states have all changed to nan.
Hello, I am a graduate student and using video inference.
After inputting a video into the model, I found that in the 18th encoder layer of the MplugOwlVisionEncoderLayer, during mlp, after QuickGELU activation function, the values of hidden_states have all changed to nan.
Hope to receive some help here. Thank you!