The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install standard machine learning libraries and NVIDIA-specific tools. All referenced resources originate from official or well-known sources.
Fetches TensorRT containers from the official NVIDIA Container Registry (nvcr.io).
References official NVIDIA documentation and repositories on GitHub (github.com/NVIDIA/TensorRT).
Directs users to the official NVIDIA developer portal for TensorRT downloads.
Installs widely used packages from the official Python Package Index (PyPI), including polygraphy, onnx, and optimum-onnx.
[INDIRECT_PROMPT_INJECTION]: The skill involves processing external data in the form of Hugging Face model IDs and ONNX files, which represents a potential ingestion surface for indirect instructions.
Ingestion points: Model IDs and ONNX file paths processed by optimum-cli and trtexec as described in SKILL.md.
Boundary markers: No specific boundary markers are used for the external IDs, though this is typical for CLI tool arguments.
Capability inventory: Includes model export, graph simplification, and engine compilation using trtexec and polygraphy.
Sanitization: The workflow incorporates onnxsim and polygraphy surgeon for graph sanitization and constant folding, which helps normalize the ingested model structure.
[COMMAND_EXECUTION]: The skill uses standard CLI tools (nvidia-smi, trtexec, polygraphy) for its primary purpose of model optimization and verification. These commands are executed locally within the user's environment to facilitate GPU-accelerated inference builds.

trt-onnx-quickstart