Vincent Nguyen

@ddaspit where did you get the information "after further research" ? When I read the mBart PAper, indeed, the Lang Token is suffixed after the source sequence and the the...

One slight correction: [eos, tgt_lang, ..., eos] on target side at training. Since I implemented NLLB-200 support in OpenNMT-py, I can confirm that prefixing instead of suffixing improves BLEU scores...

If your model's dimension is evenly divisible by the head count, the assertion model_dim % head_count == 0 should not cause an error regardless of the computer you are using....