llama.cpp Releases Version b10319 with Multimodal Bug Fix
The llama.cpp project has released version b10319, which includes a minor patch addressing a bug in the mtmd component. Specifically, the update fixes an issue where the `longest_edge` parameter ignored the minimum and maximum pixel settings. This release ensures correct image preprocessing and embedding generation for multimodal models running on llama.cpp. It prevents unexpected scaling behavior when handling images of varying resolutions. The fix resolves pull request #26638, ensuring that pixel constraints are properly respected during image scaling. The release also provides pre-built binaries for various platforms, including macOS, Windows, Linux, and Android.
## BACKGROUND
llama.cpp is an open-source port of Meta's LLaMA model in C/C++, designed for efficient LLM inference on various hardware. The mtmd (multimodal) component in llama.cpp enables multimodal support by encoding images into embeddings before feeding them into the language model. Proper image scaling and pixel constraint handling are crucial for maintaining the accuracy of these visual embeddings.