Why certain parameters are in multiple places for Nemo 2.0

[paniabhisek]

I see fp8 can be passed to GPTModel as well as MegatronMixedPrecision. Same way, precision can be passed to MegatronMixedPrecision as well as Trainer.

[ashors1]

Hi, thanks for the question, and apologies for the confusion here. These duplicate parameters essentially arise because NeMo handles precision exclusively through the mixed precision plugin, while PyTorch lightning and Megatron have slightly different ways of handling precision. Any precision-related settings in NeMo 2.0 should be specified from MegatronMixedPrecision directly. You can find some examples of various mixed precision plugins here.

[paniabhisek]

thank you so much