Skip to content

Is the decode.json generated from DeepSeek V3 ? #16

@ljliu

Description

@ljliu

there are only 120 flash kernel calls, and 118 moe kernel calls, similar to the DeepSeek V2 layers. But the hidden size is 7168 which is the same as V3 model

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions