Gemma 4 Image Crashes in Llama.cpp: The Ubatch Configuration Fix

Developers implementing Gemma 4's vision features in Llama.cpp often encounter crashes due to non-causal attention constraints that require careful ubatch configuration. The critical insight is that image token budgets (ranging from 70 to 1120 tokens depending on the task) must align perfectly with ubatch settings, making this optimization essential for stable multimodal AI deployment.