Skip to content

Perplexity Discrepancy Between Slow and Fast Paths #716

Open
@arman-kazemi

Description

@arman-kazemi

Hi,
I noticed that if the fast-path is disabled for Mamba2, the perplexity of the model increases. I tested a 130M Mamba2 model and the PPL for wikitext measured with eleuther-eval-harness is 31 when using the fast-path and ~38 when not using the fast-path (torch_forward). Do you know why this happens?

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