Skip to content

Tokenizer & Multimerge Bugs + Miscellaneous Changes #548

Open
@kromeurus

Description

@kromeurus

Hey, I found a few errors with the latest updates to mergekit plus some noticeable changes that may or may not be bugs. These errors started to show after branch #534 was pushed to main and every version since, they've persisted.

Tokenizer

If you try to do a simple passthrough merge to apply a LoRA onto a model, the model will merge but the tokenizer will fail to transfer. See image below.

Image

Strangely, if you manualy call for the LoRA's tokenizer a source for a seperate merge, it'll work. It seems to be only for a straight LoRA application merges that break.

Multimerge

After completing a merge using the !mergekit-multi command, there's a chance that mergekit will try to download all the input models all over again. This bug doesn't happen often, but it's never happened before, so I thought it might be worth at least mentioning. If the error occurs again, I'll post a screenshot of the outputs here.

Miscellaneous Changes

Ever since branch #534 was committed, I've noticed that merges take longer to complete than it's previous versions. The time deviation is minimal with smaller merges. However, larger merges are significantly slower. I tried running the same config — one pre, one post commit — and post took roughly double the time even when using --multigpu, yet the resulting merge is the same. I'm not 100% sure what causing it but I did notice the # of it/s is lower post commit then pre. Something is eating up more processing power?

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