The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to install several machine learning and benchmarking packages including vllm, torch, transformers, locust, autoawq, auto-gptq, and flash-attn from established package registries. It also references fetching pre-trained models from trusted organizations on Hugging Face, such as meta-llama and TinyLlama.
[DYNAMIC_EXECUTION]: The documentation describes the use of the --trust-remote-code flag when serving models. This feature allows vLLM to execute custom Python code provided within model repositories (e.g., for specialized model architectures). While a standard practice in the ML ecosystem, it represents a dynamic execution surface that requires the model source to be verified.
[INDIRECT_PROMPT_INJECTION]: The skill sets up an inference server that processes arbitrary text prompts, which is a known attack surface for indirect prompt injection.
Ingestion points: Untrusted data is ingested via an OpenAI-compatible API server and from batch input files like prompts.txt (SKILL.md).
Boundary markers: Present; the server configuration uses SamplingParams with defined stop tokens to manage the termination of generated output.
Capability inventory: The skill enables model inference, serving, and system monitoring functionalities (SKILL.md).
Sanitization: Absent; no specific prompt sanitization or safety filtering logic is included in the deployment instructions beyond the model's native guardrails.
[PRIVILEGE_ESCALATION]: The troubleshooting guide suggests using sudo ufw allow 8000 to resolve connectivity issues. This involves using administrative privileges to modify system firewall rules, which is a standard administrative task for server configuration (references/troubleshooting.md).

serving-llms-vllm