Index out of range errors
#32
by
manas03
- opened
I am trying to do alignment training of this model using ORPO. I am able to process the inputs using the model during evaluation loop (model in eval() mode) without any issues. However when passing the same inputs through the model during training causes index out of bounds errors. I have attached the SS of the step in modeling_llava_next.py on which this happens. Does anyone know what might be the cause ?
@manas03
can you make sure that during training you haveprocessor.tokenizer.padding_side == "right" and model.padding_side=="right". The error might be caused by the different padding sides on model and tokenizer, they have to be same