The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the download of the hqq library and optimized backends like bitblas and torchao from PyPI. It also includes instructions for downloading models from HuggingFace repositories (e.g., meta-llama, mobiuslabsgmbh), which is standard for model quantization tasks.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a structural vulnerability surface common to LLM-based tools.
Ingestion points: External model weights and configurations (AutoModelForCausalLM.from_pretrained) and user-supplied prompts (tokenizer).
Boundary markers: None explicitly defined in the provided code snippets.
Capability inventory: File system writing (model.save_pretrained) and network reading (HuggingFace API access).
Sanitization: Relies on standard tokenization; no extra validation of inputs or model metadata is described.
[DYNAMIC_EXECUTION]: The skill uses torch.compile and dynamic layer replacement to optimize model performance. These are legitimate and standard practices within the PyTorch ecosystem for improving inference efficiency.

hqq-quantization