The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to parse import statements directly from model source files (modeling_*.py) using grep and subsequently install those packages via pip.
[REMOTE_CODE_EXECUTION]: Instructions include cloning the HuggingFace transformers source code from GitHub and installing it locally, as well as importing Docker containers from the Nvidia Container Registry.
[DYNAMIC_EXECUTION]: The skill frequently uses the trust_remote_code=True flag when loading HuggingFace models and configurations, which allows for the execution of arbitrary Python scripts included in the model's repository.
[DYNAMIC_EXECUTION]: Uses importlib.import_module to load Python modules based on strings constructed from the model's configuration metadata (cfg.model_type), allowing for dynamic code loading from computed paths.
[INDIRECT_PROMPT_INJECTION]: The skill creates a significant attack surface by ingesting and processing untrusted external data (model checkpoints, configuration files, and custom scripts) that are used to drive subsequent shell commands and code execution.
Ingestion points: config.json, modeling_*.py, and HuggingFace model repositories.
Boundary markers: None identified to separate untrusted model instructions from agent logic.
Capability inventory: pip install, enroot import, sbatch job submission, and arbitrary Python execution via trust_remote_code.
Sanitization: No verification or sanitization of the model's code or configuration is performed prior to execution.
[EXTERNAL_DOWNLOADS]: Fetches Docker images from nvcr.io and source code/models from HuggingFace repositories.

modelopt-ptq