Skip to content

A bug in the implementation of the top-p sampling #60

Description

@allblueJT

return sorted_logits.scatter(1, sorted_indices, sorted_logits)

Using the sorted indices to index the sorted indices does not make sense.
I think it may be
return logits.scatter(1, sorted_indices, sorted_logits)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions