[Question] fp8 amax_history setup

[tylaar]

Hey there,

This is just a question regarding about setup FP8 amax_history. I checked the documentation about this parameter, it's clear how to set this and I understand the default value is 1024. However there is little material about how tuning this value would affect training result (I mean loss, and downstream task performance on benchmark). Also this is unclear regarding about how to select the amax algorithm selection.
May i ask for some advise regarding this?

Thanks