ggerganov/llama.cpp Released b11972 with Critical Buffer Fixes
llama.cpp release b11972 fixes out-of-bounds writing issues affecting control vector layer indices and M-RoPE position loops in multimodal encoders. These updates ensure buffer sizes and write offsets are safely checked against model limits. These bug fixes prevent potential memory corruption, segmentation faults, and undefined behaviors when running inference with custom control vectors or multimodal models like Qwen-VL. Ensuring robust bounds checking improves overall stability and security across various supported hardware platforms. The control vector update passes the model layer count to the loader and uses size_t for sizing calculations to prevent integer overflow. Meanwhile, the multimodal update iterates over whole merge tiles to prevent position buffers from overflowing on unaligned grids.
## BACKGROUND
Control vectors are float arrays that modify a language model's internal representations during inference to steer behavior without retraining weights. Multimodal Rotary Position Embedding (M-RoPE) extends standard positional encodings to handle spatial and temporal structures across vision-language inputs like text, images, and video.