Skip to content

liger-kernel version mismatch #91

@PercyHayes

Description

@PercyHayes

I ran into a problem when training the qwen3-dlm model. The function interface is incompatible. In liger-kernel 0.6.2, the forward propagation of LigerFusedLinearCrossEntropyFunction has 12 parameters, but in qwen3_dlm, the function is passed 13 parameters, which actually corresponds to the parameter format of liger-kernel 0.6.3. Same problem in llada_dlm and dream_dlm.

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