Skip to content

[zero3] nn.linear gets overriden twice #952

@stas00

Description

@stas00

running deepspeed with zero.init I get:

Your linear layers are being patched with more memory efficient version. This will persit unless manually turned reset.
Your linear layers are being patched with more memory efficient version. This will persit unless manually reset.

thanks to the inconsistent message, it actually is easy to see that this happens twice in the same process - showing that nn.linear gets overridden twice, which means it might not be able to restore the original, since at least one of them backs up itself. But it should probably be cleaned up to override only if it hasn't been already overridden or something like that...

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