llama.cpp Release b12458 Adds YuE2-3B Model Support and Multimodal Fixes
Release b12458 of llama.cpp introduces official support for the YuE2-3B music generation model along with performance improvements and API fixes for multimodal tasks. It also fixes streaming issues, refines batch APIs, and optimizes backend tensor fusions. This update expands the ecosystem's capabilities beyond traditional text and vision models into advanced audio and music generation workflows. Developers can now run complex multimodal models locally with better speed and reliability. The release adds Classifier-Free Guidance (CFG) controls and fixes logit sampling across multiple server slots to prevent sequence mixing. It also expands tensor reshape chains to allow proper backend fusion, similar to existing handling in qwen3tts.
## BACKGROUND
llama.cpp is an open-source library that enables efficient inference of large language models on consumer hardware by utilizing the GGML tensor library. Logits refer to the raw, unnormalized prediction scores outputted by a model before applying functions like softmax to determine final probabilities.