The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions and associated deployment scripts contain patterns where the output of network requests (targeting localhost) is piped directly into a Python interpreter. This occurs in SKILL.md and scripts/deploy.sh for tasks such as model identification and status verification. While targeting local services, this execution pattern presents a risk if the service output is manipulated.
[DYNAMIC_EXECUTION]: The reference file references/unsupported-models.md explicitly guides the agent to perform dynamic patching of framework source code. It instructs the agent to locate library files within system directories (e.g., /usr/local/lib) and apply modifications using sed -i or Python scripts to resolve architecture mappings or weight mismatches. This involves self-modifying or library-modifying behavior at runtime.
[EXTERNAL_DOWNLOADS]: The skill references and installs several external dependencies from established sources. It fetches Python packages from PyPI and pulls container images from NVIDIA's NGC registry and official repositories on Docker Hub for vLLM and SGLang. These are documented as standard setup procedures for the deployment environment.
[COMMAND_EXECUTION]: The skill makes extensive use of shell command execution for server lifecycle management, including the use of nohup for background processes, process group termination via kill, and environment detection via nvidia-smi and other CLI tools.

modelopt-deployment