Packages

llama.cpp is a single-repo project, but it groups its code into a few logical packages that compile and ship as separate artifacts. This section documents those packages from the perspective of a downstream consumer.

Pages

• common — common/ shared CLI helpers and Jinja chat templating.
• ggml — ggml/ tensor library and backends.
• gguf-py — Python utilities for reading/writing GGUF.

What ships out of this repo

Artifact	CMake target / file	Purpose
libllama (static or shared)	llama (src/)	C library; main consumer-facing dependency
libggml + per-backend libraries	ggml (ggml/)	Tensor library; can be reused independently
libcommon (in-tree only)	common (common/)	CLI / chat helpers; not exported by llama-config.cmake
llama-* binaries	tools/<tool>/	All the CLI programs
gguf Python package	gguf-py/	Pip-installable GGUF reader/writer
Conversion scripts	convert_*.py	HF → GGUF conversion
find_package(llama) config	cmake/llama-config.cmake.in	Downstream CMake integration

Downstream integration

find_package(llama) exposes both llama and ggml targets. The minimal example lives at examples/simple-cmake-pkg/. Header-only include/llama-cpp.h is the recommended C++ entry point.

Packaging shims live under cmake/ (llama-config.cmake.in, llama.pc.in) and the Linux/macOS/Windows installers are produced by .github/workflows/release.yml.

Versioning

There is no semver. Releases are git tags (b<number>) cut from master. ABI breakage is tracked in issue 9289 (libllama) and issue 9291 (server REST API).

License

MIT for the project as a whole. Vendored third-party code under vendor/ retains its own licenses, recorded in licenses/.