The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the llama.cpp source code from its official GitHub repository and downloads models from Hugging Face using standard CLI tools. These resources originate from reputable organizations widely recognized in the AI development community.
[COMMAND_EXECUTION]: Instructions include building software from source using make, running quantization utilities (llama-quantize, llama-imatrix), and executing inference commands via llama-cli or llama-server. These are standard procedures for the described model optimization workflow.
[PRIVILEGE_ESCALATION]: The troubleshooting documentation suggests using sudo apt install nvidia-cuda-toolkit to resolve build errors related to missing system dependencies. This is a standard administrative task for setting up a development environment.
[DYNAMIC_EXECUTION]: Python examples demonstrate the use of subprocess.run to call nvidia-smi for querying available VRAM. This diagnostic information is used to dynamically calculate the optimal number of GPU layers to offload, which is a benign hardware optimization practice.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill processes external model files in GGUF format and text files (calibration.txt, wiki.txt) used for quantization calibration or inference prompts.
Boundary markers: None specific to the data processing commands, as the workflow is intended for manual development use.
Capability inventory: Includes subprocess execution for compilation and model processing, as well as local file write operations (cat > Modelfile).
Sanitization: Standard for development-oriented documentation; inputs are processed by established CLI tools (llama-cli, llama-quantize).

gguf-quantization