Imported from SleepCloudMX/GPU-Monitor (AGENTS.md). Install upstream with npx skills add SleepCloudMX/GPU-Monitor. Copyright stays with the author.

GPU Monitor Project Rules

Runtime

• Default development environment: conda activate ai.
• Target Python 3.11 or newer.
• Keep the long-running monitor and launcher lightweight.
• Access NVIDIA NVML through nvidia-ml-py. Do not initialize CUDA merely to collect monitoring data.
• Common third-party packages are allowed. Installing missing packages into the ai environment is allowed. Ask before uninstalling packages or changing installed package versions.

Validation

Run these checks after code changes:

python -m pytest
python -m compileall src

When running on a Linux host with NVIDIA drivers, also run:

gpu-monitor --once
gpu-launcher --once ./examples/echo_hello.sh

Data

• Runtime database, locks, and logs belong under var/.
• Store timestamps in SQLite as UTC ISO 8601 strings.
• Use GPU UUIDs as stable identifiers. GPU indices are display and launch details only.