skulk-llama-server-vulkan

The pip-installable Vulkan llama-server build for Skulk's served GGUF engine (Linux x86_64).

The wheel carries the Foxlight-built llama-server and ggml-rpc-server binaries, compiled from the pinned upstream llama.cpp release with the Vulkan backend, plus the Khronos Vulkan loader (Apache-2.0) bundled alongside, so it has no Python dependencies at all. The system's one prerequisite is the GPU driver's Vulkan ICD: mesa-vulkan-drivers on AMD, the NVIDIA driver's ICD on bare-metal NVIDIA.

uv pip install skulk-llama-server-vulkan
llama-server-vulkan --list-devices

Skulk's engine provisioning discovers the installed wheel automatically and wires it as the node's served engine. Version scheme: 0.<llama.cpp build>.<packaging revision>, in lockstep with the CUDA sibling wheel and Skulk's engine pin. Built from source in the Skulk repository's engine-wheel workflow and published via PyPI trusted publishing.

This wheel is the engine path for AMD GPU nodes and the fallback for bare-metal NVIDIA nodes when the CUDA wheel is unavailable; how it fits into Skulk's install and provisioning flow is documented in the Build & Runtime Paths guide.

The bundled binaries derive from llama.cpp (MIT) and the bundled libvulkan is the Khronos Vulkan Loader (Apache-2.0); both license texts ship in the wheel under skulk_llama_server_vulkan/licenses/.