pi-vram-extension

Footer widget for the pi coding agent that shows how much GPU VRAM is in use, so you can see at a glance what your local models are eating while you work.

If you run local LLMs (llama.cpp, vLLM, LM Studio, Ollama, …) on the same machine as pi, this tells you exactly how full your card is:

VRAM 10.8GB/16.3GB

• Refreshes every 5 seconds via nvidia-smi (NVIDIA only)
• Multi-GPU systems show one reading per GPU, space separated
• Color changes with load on the hottest GPU: normal below 80%, warning at 80%+, red at 95%+ — so you know when your model is about to OOM
• If nvidia-smi is missing or fails (no NVIDIA GPU), the status line simply hides — no errors shown
• Zero runtime dependencies

Install

pi install npm:pi-vram-extension          # from npm (once published)
pi install git:github.com/LeonardBeu/pi-vram@main   # straight from this repo

Or locally from a checkout:

pi install /absolute/path/to/this-folder
# temporary, one run only:
pi -e /absolute/path/to/this-folder

After installing, the reading appears in pi's footer row next to other status items. Run /reload in an open session or start pi fresh after first install.

Notes

• Requires an NVIDIA GPU with nvidia-smi on PATH (Windows and Linux both work). AMD/Intel GPUs are not supported — nvidia-smi only sees NVIDIA cards.
• "Used" is total VRAM reserved by everything on the card: your model weights + KV cache, plus any other process (browsers, games, etc.). If you load a 14B Q4 model and see ~9GB used with nothing else running, that's roughly your footprint.
• Each refresh spawns a short-lived nvidia-smi process (~every 5s) — cheap, but not free.