The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation instructs users to install various Python packages from standard registries, including hqq, torch, transformers, vllm, bitblas, and torchao. It also includes instructions for downloading datasets from HuggingFace via the datasets library for perplexity measurement.
[COMMAND_EXECUTION]: The troubleshooting and installation sections reference system commands such as nvidia-smi and nvcc --version to verify hardware capabilities and driver versions. It also includes standard package management commands like pip install.
[DYNAMIC_EXECUTION]: The skill demonstrates the use of torch.compile for model optimization. It also showcases dynamic modification of model layers using Python's setattr to replace standard linear layers with quantized HQQLinear versions, which is a standard procedure in quantization frameworks.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of external models and datasets, which introduces a surface for indirect prompt injection if the ingested data is maliciously crafted.
Ingestion points: External model weights and configuration are loaded from the HuggingFace Hub; calibration datasets (e.g., wikitext) are loaded for sensitivity measurement and benchmarking.
Boundary markers: No explicit boundary markers or "ignore instructions" delimiters are implemented in the provided prompt templates or code snippets.
Capability inventory: The skill enables file writing (model.save_pretrained), network communication for model sharing (model.push_to_hub), and execution of system commands for environment verification.
Sanitization: The skill does not describe specific sanitization or validation routines for the content of downloaded models or datasets before processing.

hqq-quantization