The Agent Skills Directory

[COMMAND_EXECUTION]: Several helper scripts, including run_sglang_torch_profile_host.sh and run_trtllm_pytorch_profile_host.sh, orchestrate performance profiling by executing complex shell commands via docker run and docker exec. These scripts interpolate user-provided variables, such as model paths and additional server arguments (--server-extra), directly into bash execution strings (bash -lc), which can facilitate command injection if the inputs are not carefully sanitized.
[DYNAMIC_EXECUTION]: The skill includes a dedicated script, make_trtllm_py_executor_override.py, which programmatically reads, modifies, and rewrites the py_executor.py source file from the TensorRT-LLM package. This dynamically generated code is then mounted into Docker containers via volume mounts to override default execution logic, representing a sophisticated dynamic execution pattern for system-level profiling.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The analysis engine in analyze_llm_torch_profile.py and profile_common.py reads and parses torch.profiler JSON trace files and interacts with LLM server metadata and completion endpoints.
Boundary markers: The skill does not employ explicit boundary markers or specific "ignore embedded instructions" warnings when processing the contents of external trace files.
Capability inventory: The skill possesses powerful capabilities, including managing Docker containers, executing arbitrary shell commands on the host system, and performing network requests via urllib.request.
Sanitization: Trace data is loaded and processed using standard JSON parsers without content-based sanitization or validation of internal strings, which could allow maliciously crafted trace files to influence the reporting logic or downstream agent actions.
[EXTERNAL_DOWNLOADS]: The automation scripts utilize official container images from well-known providers, specifically nvcr.io/nvidia/tensorrt-llm/release:latest and vllm/vllm-openai:latest from Docker Hub.

llm-torch-profiler-analysis