llama_cpp failures: distinguish installation, native builds and model loading

Identify the failing stage

llama-cpp-python is the distribution that provides the llama_cpp import. Its documented plain pip installation may build bundled llama.cpp from source; the project also documents prebuilt CPU and selected accelerator wheels. Installation, import, native backend initialization, GGUF model load and inference are different checks. Official installation and API

python -m pip show llama-cpp-python
python -c "import sys, llama_cpp; print(sys.executable); print(llama_cpp.__file__)"

Run these with the Python executable that actually launches ComfyUI. A package record in another interpreter does not resolve No module named llama_cpp in the running instance.

Installation or native build failure

Record the exact package release, OS, architecture, intended backend and first concrete compiler/CMake error. Upstream recommends a verbose pip log for source-build failure and documents build options through CMAKE_ARGS or pip configuration settings. A Failed building wheel tail line is not itself a diagnosis. For a prebuilt wheel, compare the selected upstream index, Python/platform tags and backend requirements rather than copying a CUDA, Metal, ROCm or Vulkan command from another machine. On Windows, the upstream notes specifically discuss missing nmake or CMAKE_C_COMPILER. Change build options in an environment copy; do not rebuild the whole ComfyUI installation to investigate one package. Official installation

Import works but loading fails

First distinguish a native-library import/load error from a model error. For a GGUF model, record the exact model_path, publisher/model family, quantization, file completeness and first loader error. The upstream high-level API constructs Llama(model_path=...) before text generation; successful import therefore does not prove that a model can load or produce output. A filename ending in .gguf alone does not establish that a particular ComfyUI node or llama.cpp build supports its model architecture. Official API example

For an image-capable model, check the specific chat handler and projector or other companion files required by that model. The upstream multimodal example uses a separate clip_model_path for a LLaVA handler, while other current handlers differ; do not apply that path rule to every GGUF. A PyTorch CUDA report is not evidence that this llama.cpp backend was built or used with GPU offload. Official multimodal examples

Completion check and limits

In a copy of the ComfyUI environment, record four results separately: package installation, llama_cpp import, intended model initialization, and a small supported inference through the actual node. If GPU use matters, inspect the selected backend and runtime logs as a separate result. This site did not install the package, build the native library, load a GGUF model or run a GPU benchmark.

Original sources

• llama-cpp-python installation, backends and model API — checked 2026-09-26.

Source and English-language review: 2026-09-26. Follow the installed node and package versions rather than assuming the live upstream documentation describes an older installation.

Related checks

• Nunchaku still fails after installation: align the node, wheel and model
• Where to run ComfyUI commands: PowerShell, CMD and virtual environments