ggerganov/llama.cpp Released Version b12192
llama.cpp release b12192 introduces support for the MageVL-Cowork model along with routine optimizations for rotary positional embedding (RoPE) factors. This update ensures that model-level RoPE frequencies are properly managed and checked during initialization. Adding support for new model architectures allows developers and researchers to run the latest multi-modal models locally using efficient open-source infrastructure. This helps expand the ecosystem of models compatible with llama.cpp and its downstream deployment tools. The update refactors RoPE factors so they are shared at the clip_model level rather than being copied into individual clip_layers. The model loader also validates that v.rope_freqs is stored as F32 with dimensions matching the expected graph structure.
## BACKGROUND
llama.cpp is an open-source software library that performs inference on various large language models using the GGML tensor library. Rotary Positional Embeddings (RoPE) are a widely adopted method in neural sequence modeling for encoding absolute positions via rotations in the token embedding space.