[Daily Picks]

A LocalLLaMA thread describes a nasty silent failure: after updating llama-server, Gemma 4 multimodal features (screenshot understanding, short audio transcription) started returning garbage tokens while text chat still looked fine. No crash, model loaded normally—so it was easy to blame application code. A minimal llama-server + image/audio repro showed the same bug. Root cause (as reported): Unsloth’s Gemma 4 mmproj drifted from newer llama.cpp multimodal internals; switching the model+mmproj to ggml-org/gemma-4-E2B-it-GGUF restored vision/audio on the same build.

Practical notes: for production-ish multimodal stacks, prefer mmproj bundles that track llama.cpp (often ggml-org), or pin both the server build and the mmproj pair you tested. After any llama.cpp bump, re-test one image and one short audio clip—not just text. Keep a safety net that strips broken multimodal garbage tokens if you must stay on third-party quants temporarily. Fewer official quant options can be worth the stability.

Have you hit mmproj skew with Unsloth/bartowski (or others) after a llama.cpp update? Which combo do you pin in scripts?

Source: https://www.reddit.com/r/LocalLLaMA/comments/1vh5drv/unsloths_gemma_4_mmproj_silently_broke_vision/