llama.cpp Release b11327 Fixes Image Capping for Non-Causal Models
llama.cpp release b11327 introduces a minor fix in the multimodal (`mtmd`) module that caps `max_image` to the `ubatch` size for non-causal models. This update resolves PR #29773 to ensure proper handling of image inputs during inference. Although this is a routine patch release, it improves execution stability for developers working with multimodal and non-causal AI architectures. It prevents memory allocation errors or unexpected behavior when processing visual context alongside text. The fix enforces that the maximum number of image tokens processed simultaneously does not exceed `ubatch`, which represents the physical micro-batch size limit sent to hardware execution backends. Updated pre-compiled binaries have been published across all major operating systems and hardware targets, including CUDA, Vulkan, SYCL, OpenVINO, and Snapdragon.
## BACKGROUND
In `llama.cpp`, `ubatch` (micro-batch size) controls the maximum number of tokens evaluated in a single hardware compute step to manage memory overhead and maximize throughput. Unlike standard causal models that predict the next token based strictly on past context, non-causal models use bidirectional attention to process full context sequences simultaneously.