How can I ensemble Models at inference?

You can specify several models in the onmt_translate command line: -model model1_seed1 model2_seed2 Bear in mind that your models must share the same target vocabulary.