llama-crab-sys
Low-level FFI bindings to llama.cpp, ggml, gguf and mtmd.
Generated at build time with bindgen over the
bundled C headers and built with CMake. This crate is unsafe by design:
every public item is a thin extern "C" wrapper.
Most users should depend on llama-crab
instead. Reach for this crate only when you need direct access to raw
llama.cpp symbols.
Features
| Feature | Description |
|---|---|
openmp |
OpenMP CPU backend. Enabled by default. |
metal |
Apple Metal. Enabled by default on macOS aarch64. |
cuda / cuda-no-vmm |
NVIDIA CUDA backends. |
vulkan |
Vulkan backend. |
rocm |
AMD ROCm/HIP backend. |
opencl |
OpenCL backend. |
kleidiai |
KleidiAI CPU kernels for Arm mobile targets. |
mtmd |
Multimodal (vision + audio) helpers. |
common |
Builds libcommon.a for chat templates and JSON schema helpers. |
llguidance |
llguidance sampler integration. |
dynamic-link |
Link against libllama as a shared object. |
system-ggml |
Use a system GGML installation. |
dynamic-backends |
Load GGML backends at runtime. |
Resources
License
Licensed under the MIT License.