The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads resources from several remote locations including NVIDIA's official GitHub repositories, the NVIDIA Container Registry (nvcr.io), and the HuggingFace Hub. \n
Clones tao-core, tao-pytorch, tao-deploy, and tao-dataservices from NVIDIA organizations. \n
Pulls official TAO Toolkit container images from nvcr.io. \n
Downloads model weights and configurations from huggingface.co using the transformers library.\n- [COMMAND_EXECUTION]: Extensive use of shell commands and Docker operations to manage the integration lifecycle. \n
Uses docker run and docker exec to isolate development tasks inside containers. \n
Performs source installations using pip install and setup.py develop within these containers. \n
Uses git for repository and branch management.\n- [DYNAMIC_EXECUTION]: Runtime generation and execution of Python code are used to perform model inspection and validation. \n
Dynamically executes Python probes via docker exec to verify model compatibility and extract metadata. \n
Generates implementation code (e.g., adapters and trainers) that is executed as part of the TAO pipeline.\n- [CREDENTIALS_UNSAFE]: The skill manages sensitive tokens for HuggingFace and NGC. \n
Recommends the use of environment variables (HF_TOKEN, NGC_KEY) and secure docker login practices. \n
No hardcoded secrets were found in the skill content.\n- [INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection by processing external data from the HuggingFace Hub. \n
Ingestion points: Reads config.json and model weights from user-specified HuggingFace Model IDs. \n
Boundary markers: No explicit boundaries are placed around the fetched model metadata. \n
Capability inventory: The skill has high capabilities, including full filesystem access (via Docker bind-mounts), network operations, and arbitrary command execution. \n
Sanitization: There is no explicit sanitization of model configuration data before it is used to drive the implementation logic.

tao-port-huggingface-model