Skip to content

stage 2 training with CP>1 #137

@berserkr

Description

@berserkr

Hi, it is my understanding that calculate_per_token_loss: false is required for sample-level loss... Stage 2 in the paper, however, it seems that this requires CP==1? Does this mean that Nemotron at 512k was trained without CP?

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions