Skip to content

about mtp speculative decoding #35

@teie-s

Description

@teie-s

Hi, thanks for releasing MiMo!
I have a simple question about the architecture:
If one MTP layer can already generate a multi-token draft sequence, what are the other MTP layers used for?
In the paper you mention 3× MTP layers.
But conceptually, a single MTP block already produces multiple future tokens.
So I want to confirm:
Do the extra MTP layers predict different future positions?
Or do they contribute to the same draft sequence?
What benefit do layers 2 and 3 provide compared to using just one MTP layer?
Thanks!

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