ggerganov/llama.cpp Released Version b12021 with Gemma 4 Updates
Release b12021 of llama.cpp introduces multi-modal updates, specifically improving how Gemma 4 images are resized without using letterbox padding. This update improves image processing efficiency and fidelity for multi-modal models within the llama.cpp ecosystem, benefiting users running local vision-language models. The pull request #30922 modifies the multi-modal (mtmd) module to handle image scaling directly without padding, optimizing resource utilization during inference.
## BACKGROUND
llama.cpp is a popular open-source library that enables efficient inference of large language models on consumer hardware, co-developed alongside the GGML tensor library. Letterbox padding is an image processing technique used to preserve an image's aspect ratio by adding borders, which this update avoids for Gemma 4.