The Agent Skills Directory

[SAFE]: The skill follows established best practices for machine learning development. It relies on official package registries (PyPI) and well-known model hosting platforms (HuggingFace).
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates processing of external models and datasets (e.g., WikiText). While this constitutes an attack surface common to all LLM-based workflows, the implementation uses standard library functions without introducing new vulnerabilities. Any risk is mitigated by standard LLM guardrails.
[DYNAMIC_EXECUTION]: The skill references torch.compile and various CUDA kernel backends (Marlin, BitBlas, ATEN). These are standard performance optimization techniques in the PyTorch ecosystem and do not involve unsafe dynamic loading of untrusted code.

hqq-quantization