CMake Build Options

Complete reference for CMake configuration options in llama.cpp

Overview

llama.cpp provides extensive CMake options to customize your build. Options are specified with -D flag:

cmake -B build -DOPTION_NAME=VALUE

General Options

Build Type

Option	Values	Default	Description
CMAKE_BUILD_TYPE	Debug, Release, RelWithDebInfo, MinSizeRel	Release	Build configuration type
BUILD_SHARED_LIBS	ON, OFF	Platform-dependent	Build shared libraries instead of static
GGML_STATIC	ON, OFF	OFF	Static link libraries

Debug build

cmake -B build -DCMAKE_BUILD_TYPE=Debug
cmake --build build

Static build

cmake -B build -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release

Optimization Options

Option	Values	Default	Description
GGML_NATIVE	ON, OFF	ON	Optimize for current system CPU
GGML_LTO	ON, OFF	OFF	Enable link-time optimization
GGML_CCACHE	ON, OFF	ON	Use ccache if available

GGML_NATIVE=ON optimizes for your current CPU but may not work on other systems.

Portable build (no native optimization)

cmake -B build -DGGML_NATIVE=OFF
cmake --build build --config Release

Build Targets

Option	Values	Default	Description
LLAMA_BUILD_TESTS	ON, OFF	ON (standalone)	Build test suite
LLAMA_BUILD_TOOLS	ON, OFF	ON (standalone)	Build command-line tools
LLAMA_BUILD_EXAMPLES	ON, OFF	ON (standalone)	Build example programs
LLAMA_BUILD_SERVER	ON, OFF	ON (standalone)	Build HTTP server
LLAMA_BUILD_COMMON	ON, OFF	ON (standalone)	Build common utilities

Minimal build (library only)

cmake -B build \
    -DLLAMA_BUILD_TESTS=OFF \
    -DLLAMA_BUILD_TOOLS=OFF \
    -DLLAMA_BUILD_EXAMPLES=OFF \
    -DLLAMA_BUILD_SERVER=OFF
cmake --build build --config Release

GPU Backend Options

CUDA (NVIDIA)

Option	Values	Default	Description
GGML_CUDA	ON, OFF	OFF	Enable CUDA support
GGML_CUDA_FORCE_MMQ	ON, OFF	OFF	Force MMQ kernels instead of cuBLAS
GGML_CUDA_FORCE_CUBLAS	ON, OFF	OFF	Force cuBLAS instead of custom kernels
GGML_CUDA_GRAPHS	ON, OFF	ON	Enable CUDA graphs
GGML_CUDA_FA	ON, OFF	ON	Compile FlashAttention kernels
GGML_CUDA_FA_ALL_QUANTS	ON, OFF	OFF	Compile all quants for FlashAttention
GGML_CUDA_PEER_MAX_BATCH_SIZE	Integer	128	Max batch size for peer access
GGML_CUDA_NO_PEER_COPY	ON, OFF	OFF	Disable peer-to-peer copies
GGML_CUDA_NO_VMM	ON, OFF	OFF	Disable CUDA Virtual Memory Management
CMAKE_CUDA_ARCHITECTURES	String	Auto	Specify compute capabilities (e.g., “86;89”)

Basic CUDA

cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release

GGML_CUDA_FA_ALL_QUANTS=ON significantly increases compilation time and binary size.

HIP (AMD)

Option	Values	Default	Description
GGML_HIP	ON, OFF	OFF	Enable HIP support
GGML_HIP_GRAPHS	ON, OFF	OFF	Enable HIP graphs (experimental)
GGML_HIP_NO_VMM	ON, OFF	ON	Disable HIP Virtual Memory Management
GGML_HIP_ROCWMMA_FATTN	ON, OFF	OFF	Enable rocWMMA for FlashAttention
GGML_HIP_MMQ_MFMA	ON, OFF	ON	Enable MFMA MMA for CDNA in MMQ
GPU_TARGETS	String	All	Specify GPU architecture (e.g., “gfx1030”)

Linux HIP build

HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
    cmake -B build -DGGML_HIP=ON -DGPU_TARGETS=gfx1030
cmake --build build --config Release

Windows HIP build

set PATH=%HIP_PATH%\bin;%PATH%
cmake -B build -G Ninja -DGGML_HIP=ON -DGPU_TARGETS=gfx1100 \
    -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++
cmake --build build

Metal (Apple)

Option	Values	Default	Description
GGML_METAL	ON, OFF	ON (macOS)	Enable Metal GPU support
GGML_METAL_NDEBUG	ON, OFF	OFF	Disable Metal debugging
GGML_METAL_SHADER_DEBUG	ON, OFF	OFF	Compile Metal with -fno-fast-math
GGML_METAL_EMBED_LIBRARY	ON, OFF	ON	Embed Metal library in binary
GGML_METAL_MACOSX_VERSION_MIN	String	Empty	Metal minimum macOS version

Disable Metal

cmake -B build -DGGML_METAL=OFF
cmake --build build --config Release

Vulkan

Option	Values	Default	Description
GGML_VULKAN	ON, OFF	OFF	Enable Vulkan support
GGML_VULKAN_CHECK_RESULTS	ON, OFF	OFF	Run Vulkan operation checks
GGML_VULKAN_DEBUG	ON, OFF	OFF	Enable Vulkan debug output
GGML_VULKAN_MEMORY_DEBUG	ON, OFF	OFF	Enable Vulkan memory debug
GGML_VULKAN_VALIDATE	ON, OFF	OFF	Enable Vulkan validation layers

Vulkan with validation

cmake -B build -DGGML_VULKAN=ON -DGGML_VULKAN_VALIDATE=ON
cmake --build build --config Release

SYCL (Intel GPU)

Option	Values	Default	Description
GGML_SYCL	ON, OFF	OFF	Enable SYCL support
GGML_SYCL_F16	ON, OFF	OFF	Use 16-bit floats for calculations
GGML_SYCL_GRAPH	ON, OFF	ON	Enable graphs in SYCL backend
GGML_SYCL_DNN	ON, OFF	ON	Enable oneDNN in SYCL backend
GGML_SYCL_TARGET	String	INTEL	SYCL target device

cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_F16=ON
cmake --build build --config Release

CPU Backend Options

BLAS Libraries

Option	Values	Default	Description
GGML_BLAS	ON, OFF	OFF	Enable BLAS support
GGML_BLAS_VENDOR	String	Generic	BLAS vendor (OpenBLAS, Intel10_64lp, Apple, etc.)
GGML_ACCELERATE	ON, OFF	ON (macOS)	Use Accelerate framework

OpenBLAS

cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
cmake --build build --config Release

CPU Instruction Sets

Option	Values	Default	Description
GGML_SSE42	ON, OFF	ON*	Enable SSE 4.2
GGML_AVX	ON, OFF	ON*	Enable AVX
GGML_AVX2	ON, OFF	ON*	Enable AVX2
GGML_AVX_VNNI	ON, OFF	OFF	Enable AVX-VNNI
GGML_AVX512	ON, OFF	OFF	Enable AVX-512F
GGML_AVX512_VBMI	ON, OFF	OFF	Enable AVX-512 VBMI
GGML_AVX512_VNNI	ON, OFF	OFF	Enable AVX-512 VNNI
GGML_AVX512_BF16	ON, OFF	OFF	Enable AVX-512 BF16
GGML_FMA	ON, OFF	ON*	Enable FMA (non-MSVC)
GGML_F16C	ON, OFF	ON*	Enable F16C (non-MSVC)
GGML_BMI2	ON, OFF	ON*	Enable BMI2

* Default is ON unless GGML_NATIVE=OFF

These options override automatic detection. Only use if you know your target CPU capabilities.

Enable AVX-512

cmake -B build -DGGML_AVX512=ON -DGGML_AVX512_VNNI=ON
cmake --build build --config Release

ARM CPU Options

Option	Values	Default	Description
GGML_CPU_KLEIDIAI	ON, OFF	OFF	Use KleidiAI optimized kernels
GGML_CPU_ARM_ARCH	String	Empty	CPU architecture for ARM
GGML_RVV	ON, OFF	ON	Enable RISC-V vector extension

Arm KleidiAI

cmake -B build -DGGML_CPU_KLEIDIAI=ON
cmake --build build --config Release

Special Purpose Backends

CANN (Ascend NPU)

cmake -B build -DGGML_CANN=ON
cmake --build build --config Release

ZenDNN (AMD EPYC)

Option	Values	Default	Description
GGML_ZENDNN	ON, OFF	OFF	Enable ZenDNN support
ZENDNN_ROOT	Path	Empty	Path to ZenDNN installation

cmake -B build -DGGML_ZENDNN=ON
cmake --build build --config Release

OpenCL (Qualcomm Adreno)

Option	Values	Default	Description
GGML_OPENCL	ON, OFF	OFF	Enable OpenCL support
GGML_OPENCL_PROFILING	ON, OFF	OFF	Enable OpenCL profiling
GGML_OPENCL_EMBED_KERNELS	ON, OFF	ON	Embed kernels in binary
GGML_OPENCL_USE_ADRENO_KERNELS	ON, OFF	ON	Use optimized Adreno kernels

cmake -B build -DGGML_OPENCL=ON
cmake --build build --config Release

RPC Backend

cmake -B build -DGGML_RPC=ON
cmake --build build --config Release

Advanced Options

Threading

Option	Values	Default	Description
GGML_OPENMP	ON, OFF	ON	Enable OpenMP for threading

Disable OpenMP

cmake -B build -DGGML_OPENMP=OFF
cmake --build build --config Release

Debugging & Diagnostics

Option	Values	Default	Description
LLAMA_ALL_WARNINGS	ON, OFF	ON	Enable all compiler warnings
LLAMA_FATAL_WARNINGS	ON, OFF	OFF	Treat warnings as errors
LLAMA_SANITIZE_THREAD	ON, OFF	OFF	Enable thread sanitizer
LLAMA_SANITIZE_ADDRESS	ON, OFF	OFF	Enable address sanitizer
LLAMA_SANITIZE_UNDEFINED	ON, OFF	OFF	Enable undefined behavior sanitizer

Build with sanitizers

cmake -B build -DCMAKE_BUILD_TYPE=Debug \
    -DLLAMA_SANITIZE_ADDRESS=ON \
    -DLLAMA_SANITIZE_UNDEFINED=ON
cmake --build build

Backend Dynamic Loading

Option	Values	Default	Description
GGML_BACKEND_DL	ON, OFF	OFF	Build backends as dynamic libraries
GGML_BACKEND_DIR	Path	Empty	Directory to load backends from

GGML_BACKEND_DL requires BUILD_SHARED_LIBS=ON.

Dynamic backend loading

cmake -B build -DBUILD_SHARED_LIBS=ON -DGGML_BACKEND_DL=ON
cmake --build build --config Release

HTTPS/SSL Support

Option	Values	Default	Description
LLAMA_OPENSSL	ON, OFF	ON	Use OpenSSL for HTTPS support

Disable OpenSSL

cmake -B build -DLLAMA_OPENSSL=OFF
cmake --build build --config Release

Common Build Configurations

Development Build

Fast compilation, full debug info:

cmake -B build -DCMAKE_BUILD_TYPE=Debug \
    -DLLAMA_ALL_WARNINGS=ON
cmake --build build

Production Build

Optimized for performance:

cmake -B build -DCMAKE_BUILD_TYPE=Release \
    -DGGML_NATIVE=ON \
    -DGGML_LTO=ON \
    -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release

Multi-GPU Build

CUDA + Vulkan for flexibility:

cmake -B build -DGGML_CUDA=ON -DGGML_VULKAN=ON \
    -DGGML_CUDA_GRAPHS=ON
cmake --build build --config Release

Minimal Binary Size

Library only, static linking:

cmake -B build -DCMAKE_BUILD_TYPE=MinSizeRel \
    -DBUILD_SHARED_LIBS=OFF \
    -DLLAMA_BUILD_TESTS=OFF \
    -DLLAMA_BUILD_TOOLS=OFF \
    -DLLAMA_BUILD_EXAMPLES=OFF \
    -DLLAMA_BUILD_SERVER=OFF
cmake --build build --config MinSizeRel

Portable Build

Runs on any x86_64 system:

cmake -B build -DGGML_NATIVE=OFF \
    -DGGML_AVX=ON \
    -DGGML_AVX2=ON \
    -DGGML_FMA=ON \
    -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release

Option Discovery

To see all available CMake options:

cmake -B build -LH

To see current configuration:

cmake -B build -LA

For GUI-based configuration (requires cmake-gui):

cmake-gui -B build