llama.cpp Release b10771 Refactors Multimodal Tokenization API
llama.cpp release b10771 introduces the `mtmd_tokenize_from_parts()` function to improve tokenization handling in multimodal workflows. The update modifies the `mtmd-cli` tool to use this new API and shifts the `add_special` token configuration to the call level. Although this is a minor maintenance release, it refines internal API design for processing structured prompts made of multiple content types like text, images, or audio. Developers extending llama.cpp's multimodal features gain cleaner control over token generation and special token handling. Implemented in PR #28250, `mtmd_tokenize_from_parts()` allows tokenizing separate parts of a prompt in sequence. Moving `add_special` to the call level provides callers finer granularity over whether special tokens are appended during individual tokenization passes.
## BACKGROUND
llama.cpp is a widely used open-source C/C++ framework for running LLM inference efficiently across various CPU and GPU architectures. In the llama.cpp ecosystem, `mtmd` refers to the multimodal CLI tools designed to process mixed inputs such as text, images, and audio before sending them to a model. Tokenization is the step that breaks down these inputs into numerical token IDs suitable for model consumption.