decoder_input_ids must be set to 2 when using fairseq model

[cocovoc]

Issue

When using fairseq and TensorRT-LLM for inference, I encountered an issue.
In the model's vocabulary,：

</s>:       0
<pad>:    2
</s>:       0
<unk>:    1

When using TensorRT-LLM, the decoder_input_ids must be set to 2 (the token) in order to function correctly.
If I set decoder_input_ids to other token IDs (e.g., 0 for ““ or 1 for “”), the model does not work properly and does not produce the expected output.

Environment

fairseq version: 0.12.2
TensorRT-LLM version: 0.9.0

Example

the output of the fairseq model :

the output of the TensorRT-LLM (decoder_start_token_id =2):

the output of the TensorRT-LLM (decoder_start_token_id =0):

Does the token_id in the TensorRT-LLM engine need to be consistent with the token_id in the fairseq model?