Skip to content

mtmd: fix gemma 4 projector pre_norm#23822

Merged
ngxson merged 1 commit into
masterfrom
xsn/fix_gemma4_prenorm
May 28, 2026
Merged

mtmd: fix gemma 4 projector pre_norm#23822
ngxson merged 1 commit into
masterfrom
xsn/fix_gemma4_prenorm

Conversation

@ngxson
Copy link
Copy Markdown
Contributor

@ngxson ngxson commented May 28, 2026

Overview

The early pre-release version of gemma 4 uses post norm for multimodal projector, but it was later swapped to pre-norm and I did not notice about that.

This should fix some vision-related problems with gemma 4.

Ref python code: https://github.com/huggingface/transformers/blob/1656d90b774d94c30af24113e60e926fc2f39072/src/transformers/models/gemma4/modeling_gemma4.py#L2068-L2092

Test:

image image image

Requirements

@ngxson ngxson requested a review from a team as a code owner May 28, 2026 14:42
@ngxson
Copy link
Copy Markdown
Contributor Author

ngxson commented May 28, 2026

asking for 2nd approval @ggml-org/maintainers 🙏

@ngxson ngxson merged commit c8914ad into master May 28, 2026
24 of 27 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants