sbakus1528/gemma-4-E4B-it-mmproj-vision

gemma-4-E4B-it mmproj — vision only

The official Gemma 4 E4B multimodal projector ships vision and audio in one file. If your app only needs image understanding, the audio encoder is dead weight — and llama.cpp has no flag to disable one modality (--no-mmproj turns off both).

This is the same projector with the audio encoder removed. No weight value was modified.

On an 8 GB laptop GPU that is roughly one extra slot.

Verified, not assumed

Both files were loaded in llama-server and asked the same questions with the same image:

• —synthetic image (3 blue squares, "how many?") — original: 3, this: 3
• —real 1920x1080 desktop screenshot ("what is this screen?") — both correctly identified Google Colab and that machine-learning work was in progress

A first attempt loaded fine but returned garbage (<unused49>): the repacking had flattened clip.vision.image_mean / image_std from [0,0,0] / [1,1,1] to scalars, so image normalization was wrong. Every key and tensor was present — only the values were broken. Worth knowing if you repack one yourself: comparing key lists is not enough, compare values.

What changed

• —removed tensors: a.* (751) and mm.a.input_projection.weight
• —kept tensors: v.* (658) and mm.input_projection.weight
• —metadata: clip.has_audio_encoder = false; clip.audio.* keys removed

Usage

llama-server -m gemma-4-E4B-it-Q4_K_M.gguf \
             --mmproj gemma-4-E4B-it-mmproj-vision.gguf -ngl 99

Audio input will not work with this file — that is the point. Use the original projector if you need it.

License

Apache-2.0, inherited from google/gemma-4-E4B-it. Source of the GGUF conversion: unsloth/gemma-4-E4B-it-GGUF. See NOTICE for the required attribution and the list of changes.