Use decoder_input_ids with deepspeed

This is a question about how to use decoder_input_ids with deepspeed.
I have a language generation task that has decoder_input_ids. For example, I need the generation model to start the generation with ‘yes’.
My solution is that I pass the decoder_input_ids to model.generate() by inputs['decoder_input_ids'] within Seq2SeqTrainer, is this the right way?

Edit: It doesn’t seem to work that way. I suspect that the does not parse the decoder_input_ids as an argument inside the model(*input). Any suggestions?