llm-torch-profiler-analysis
Warn
Audited by Gen Agent Trust Hub on Sep 16, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill uses multiple bash scripts (
run_sglang_torch_profile_host.sh,run_vllm_torch_profile_host.sh, etc.) to execute host-level shell commands. These commands includedocker run,docker exec, andpkill, which are used to manage LLM server containers and process execution traces. - [DYNAMIC_EXECUTION]: Several components perform dynamic execution. The utility
make_trtllm_py_executor_override.pyreads a Python source file from a Docker image and performs string-based patching to inject new parameters at runtime. Additionally, shell scripts use Python heredocs to execute inline code for processing server responses and timeouts. - [EXTERNAL_DOWNLOADS]: The skill pulls Docker images from the official NVIDIA Container Registry (
nvcr.io) and interacts with Hugging Face to download model weights. While these are well-known and trusted services, the skill requires validHF_TOKENcredentials to be passed through environment variables to the containers.
Audit Metadata