Hi,
I am using a T5ForConditionalGeneration model for its decoder while using an outside encoder. I access only the decoder by using the model forward call by model(encoder_outputs=encoutputs, labels=label)
I want to generate new outputs without teacher forcing, and the vanilla way to do this appear to be model.generate(), but model.generate() only allows for “input_ids”, not “encoder_outputs”. how can I use “encoder_outputs” for generation?