Johann Frei

**Describe the bug** This issue refers to the code line at: https://github.com/EleutherAI/gpt-neox/blob/1cee5b7c7074302de4867ad5cac3f1ea26f7a7d7/megatron/text_generation_utils.py#L100C43-L100C50 To my understanding, the top-p should be applied on the pre-top-k-filtered token probabilities. Apparently though, if top-k and...