Dependencies

llama.cpp is intentionally lean. The CONTRIBUTING guide tells contributors to "avoid adding third-party dependencies." Most "dependencies" are vendored single-header libraries under vendor/; only a handful are real, system-level packages, and they are all gated by build flags.

System (build-time, optional)

Dependency	Required when	Notes
CMake ≥ 3.14	Always	Build system
C++17 compiler	Always	gcc / clang / MSVC / Apple clang
Python ≥ 3.9	Building Python tooling	convert_*.py, gguf-py
CUDA Toolkit	-DGGML_CUDA=ON	NVIDIA backend
HIP / ROCm	-DGGML_HIP=ON	AMD backend
Metal SDK	macOS by default	Apple backend
Vulkan SDK	-DGGML_VULKAN=ON	glslc for shader compilation
Intel oneAPI / DPC++	-DGGML_SYCL=ON	SYCL backend
MUSA SDK	-DGGML_MUSA=ON	Moore Threads
OpenCL	-DGGML_OPENCL=ON	Adreno path
Qualcomm Hexagon SDK	-DGGML_HEXAGON=ON	DSP path
Huawei CANN	-DGGML_CANN=ON	Ascend NPU
OpenVINO Runtime	-DGGML_OPENVINO=ON	Intel CPU/iGPU/NPU
WebGPU runtime (Dawn / wgpu-native)	-DGGML_WEBGPU=ON	WebGPU
OpenMP	-DGGML_OPENMP=ON	Optional CPU threading
BLAS (OpenBLAS / MKL / Accelerate)	-DGGML_BLAS=ON	Prompt-processing matmul
libcurl	-DLLAMA_CURL=ON	-hf <repo> downloads
OpenSSL	-DLLAMA_SERVER_SSL=ON	HTTPS server
zDNN library	-DGGML_ZDNN=ON	IBM Z
ZenDNN library	-DGGML_ZENDNN=ON	AMD Zen CPU

Without any of these, you still get a fully working CPU build.

Vendored (vendor/)

These ship in-tree as single-header or minimal-source libraries. None require external installation.

Library	Purpose
vendor/nlohmann/json.hpp	JSON parsing/printing for the server, autoparser, and tooling
vendor/cpp-httplib/httplib.h	Single-header HTTP server used by tools/server and tools/rpc
vendor/minja/minja.hpp	Jinja2-compatible template engine used by common/chat.cpp
vendor/stb/stb_image.h	Image loading for tools/mtmd
vendor/miniaudio/miniaudio.h	Audio I/O for tools/tts and tools/mtmd-audio
vendor/llguidance/	Optional Rust grammar engine (build-gated by LLAMA_LLGUIDANCE)
vendor/cpp-jsonschema/	JSON Schema validation helpers

Third-party licenses are recorded in licenses/.

Python (requirements*.txt, pyproject.toml)

Python tooling lives in gguf-py/ and the top-level convert_*.py scripts. Requirements are split into focused files under requirements/ and aggregated by requirements.txt at the repo root. gguf-py/pyproject.toml is poetry-managed.

Top-level packages (subset):

• numpy — tensor math during conversion.
• torch — used by some conversion paths to load HF checkpoints.
• safetensors — read modern HF checkpoints.
• sentencepiece, tokenizers — reference tokenizers used during conversion / vocab generation.
• protobuf — sentencepiece model files.
• huggingface_hub — used by some conversion helpers.

For development:

• pytest (server tests, gguf-py tests).
• flake8, mypy, pyright, ty — linters/type checkers.
• pre-commit — local hook runner.

Build artifacts

A standard Release build produces (with default flags):

• build/bin/llama-cli
• build/bin/llama-server
• build/bin/llama-quantize
• build/bin/llama-bench
• build/bin/llama-imatrix
• build/bin/llama-perplexity
• build/bin/llama-tokenize
• build/bin/llama-mtmd-cli
• build/bin/llama-gguf-split
• build/bin/test-backend-ops (with -DLLAMA_BUILD_TESTS=ON)
• ... plus example binaries and any backend plugins built as shared libraries

make install (CMake) installs libllama, libggml, headers, and the package config under ${CMAKE_INSTALL_PREFIX}.