Hugging Face's llama.cpp team released version 1.0.0 on Nov. 13, committing to clearer version changes when updates break compatibility with applications built on its software, the team said.
The team said developers increasingly depend on its interfaces, making stability more important. Hugging Face published the announcement on Nov. 13 under the title “llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem”.
The team said both llama.cpp and its underlying machine-learning library, ggml, will change their major version numbers whenever updates introduce incompatible interfaces. Those interfaces govern how applications communicate with the libraries.
The commitment covers compatibility at both the programming and compiled-software levels, the team said. It would become stricter about making incompatible changes visible to developers using its libraries as adoption grows.
The team released version 1.0.0 of both libraries. It described llama.cpp as software for running large language models in C and C++, while ggml provides the underlying mathematical operations used in machine learning.
The project has grown over the three and a half years since it began, and many applications now depend on it, the team said. It linked the versioning commitment to the growing number of developers relying on its work.
The software can run models that handle multiple types of input across hardware ranging from consumer computers to enterprise servers, the team said. Its approach combines reduced numerical precision, hardware-specific processing and portable C/C++ code.
The team said llama.cpp supports more than 150 model architecture variants, or different designs for how models process information. The software also supports all mainstream speculative decoding methods, which help models generate responses faster.
The team said ggml supports different hardware through dedicated backends, the software components that connect its calculations to particular computing platforms. Developers can compile each backend separately and choose to load it while an application runs.
The team said it was putting more resources into automated testing infrastructure to support faster development. Hugging Face's Buckets and Jobs services help expand testing capacity for changes entering the projects' main development branches, it said.
NVIDIA, AMD, Qualcomm, Arm, IBM and Intel provide hardware support, the team said. It was adding more machines to run automated checks across the major hardware platforms.
The team said it was incorporating large language models into its development work while keeping code quality, control over software architecture and understanding of implementation details among its priorities.
The team also said it wanted to expand how developers and end users obtain its software. It committed to continued support for model architectures and said it expected local AI and privacy to become increasingly important.
