Does the TPF calculation include the EOS token?

[zs1314]

Hello! I noticed in your code you compute: num_tokens += (generated_answer_ids != 126081).sum(). If your TPF is defined as num_tokens / NFE, then LLaDA’s result shouldn’t be 1—it should be less than 1. Can you explain why?

[ZinYY]

Thank you for your interest in our work and for carefully reading our code. You are correct — in our implementation, TPF is computed by re-encoding the dLLM's final response text, which excludes any EOS tokens that are generated but do not appear in the decoded output. For LLaDA, this means the true TPF is slightly below 1. We will correct this in the revised version.