After three and a half years llama.cpp reaches version 1.0, and anyone breaking the API must now show it in the major number
The library for running language models locally switches to semantic versioning and immediately drops a batch of deprecated functions.
The developers at ggml-org on Friday released version 1.0.0 of llama.cpp, the popular open-source library for running language models on your own computer, together with version 1.0.0 of the underlying tensor library ggml. Georgi Gerganov's project is about three and a half years old, Runtime Wire reports.
The most important change lies in the rules rather than the code. All ggml-org projects now follow semantic versioning: a change that breaks the public C API requires a new major number, new features or model support arrive in a minor release and bug fixes in a patch, AI TL;DR summarises. A writer on DEV Community sees it above all as a new contract with those who build on the library.
Anyone upgrading needs to take care: more than thirty deprecated functions have been removed, so code that still calls them no longer compiles. New additions include a KV cache in FP8 format, and llama-server is better protected against SSRF and DNS rebinding attacks. The library now supports more than 150 model architectures, according to DEV Community.