GPU VRAM Cleaner

Tired of not being able to kill the processes that eat up your GPU? Now you finally can.

A terminal UI (Textual + colorama) for Windows Terminal / PowerShell that shows every NVIDIA GPU in the system and the processes holding its VRAM, and lets you close them to free memory.

Features

• One panel per GPU with model, driver version, temperature (°C), load, VRAM used/total, core and memory clocks (MHz) and power draw against the power limit (W).

• Three live graphs per GPU, refreshed every second: VRAM %, fan % and load %.

• Percentages and temperatures change color with their level (green / yellow / red):

  Reading	Green	Yellow	Red
  VRAM, load	< 60 %	60–84 %	≥ 85 %
  Fan	< 50 %	50–79 %	≥ 80 %
  Temperature	< 65 °C	65–79 °C	≥ 80 °C

• Process list with each process's VRAM and GPU load % (the busiest GPU engine it uses, like Task Manager's GPU column, colored by level). Move with ↑ / ↓, press Del to get a Close process confirmation with Yes (selected by default) and Cancel (use ← / → to switch, Enter to confirm, Esc to cancel).

• Menu (m) with:

  • Change theme: 7 color themes for borders and backgrounds (NVIDIA Green, Deep Ocean, Dracula Night, Solar Flare, Nord Frost, Paper Light and Synthwave Neon, the default). Themes preview live as you move; the choice is saved.
  • Radical clean: closes every process using VRAM in one go. It is marked Dangerous: two red confirmation popups, the second one asking "Are you sure? This can be potentially unsafe for your system." (Cancel is the default in both).

Close process	Radical clean (Dangerous)

NVIDIA Green theme	Theme menu

Keeping Windows safe

Both the single close and the radical clean refuse to touch processes that Windows needs, so the desktop, login session and drivers keep working:

• core system processes (System, csrss.exe, wininit.exe, winlogon.exe, services.exe, lsass.exe, svchost.exe, dwm.exe, explorer.exe, sihost.exe, ctfmon.exe, shell hosts, NVIDIA display container, ...);
• anything running as SYSTEM, LOCAL SERVICE or NETWORK SERVICE, or that the app cannot inspect;
• the app itself and its parent processes (the PowerShell and Windows Terminal running it).

Protected rows are marked protected in the list. Everything else that holds VRAM (games, browsers, Python/CUDA jobs, Blender, OBS, Discord, ...) is closed: first politely (terminate), then forcefully (kill) if it does not exit within 3 seconds. The whole app is closed, not just the process holding VRAM: browsers and Electron apps (Chrome, Edge, Discord, Claude, VS Code) keep the GPU in a helper process that the main window would respawn at once. Closing processes owned by other users requires running the terminal as administrator.

Install

gpu-vram-cleaner is on PyPI and works on Windows and Linux with an NVIDIA GPU and its driver installed (NVML ships with the driver). On macOS, or any machine without an NVIDIA card, it runs in --demo mode.

System-wide (recommended)

Install gpu-cleaner as a command available from any terminal with uv:

uv tool install gpu-vram-cleaner
uv tool update-shell   # only the first time: adds uv's tool folder to your PATH

If you don't have uv yet, install it first:

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

# Linux / macOS
curl -LsSf https://astral.sh/uv/install.sh | sh

pipx works too: pipx install gpu-vram-cleaner.

Open a new terminal and run:

gpu-cleaner

Update to the latest version with uv tool upgrade gpu-vram-cleaner, and remove it with uv tool uninstall gpu-vram-cleaner.

Tip: to close processes owned by other users or by system services, open the terminal as administrator (Windows) or run with sudo (Linux).

From a clone (development)

git clone https://github.com/andriusland/GPU-VRAM-Cleaner
cd GPU-VRAM-Cleaner
uv run gpu-cleaner

Options:

--demo          simulated GPUs and processes (nothing real is closed); handy without an NVIDIA card
--interval S    refresh interval in seconds (default 1)
--autoclean     close every unprotected process using VRAM and exit, without the UI
--dry-run       with --autoclean: only list what would be closed

Keys: ↑/↓ select · Del close process · m menu · t theme · x radical clean · q quit.

Clean VRAM from scripts

--autoclean runs the same protected radical clean as x, with no UI and no confirmation, so you can free the GPU right before a training run:

gpu-cleaner --autoclean --dry-run          # preview what would be closed
gpu-cleaner --autoclean; python train.py   # PowerShell

gpu-cleaner --autoclean && python train.py   # Linux: only train if every process was closed

It exits with 0 when every target is gone (or there was nothing to close) and 1 if any process could not be closed (for example, owned by another user without administrator rights).

On Windows, NVML cannot report per-process VRAM under the WDDM driver model (every GeForce card), so the app reads it from the Windows "GPU Process Memory" performance counters instead, the same source as Task Manager's "Dedicated GPU memory" column. Per-process load comes from the "GPU Engine" counters the same way (it shows N/A for the first second while Windows takes its first sample).

Development

The project was built test first (red → green → refactor).

uv run pytest          # tests (core logic + Textual UI driven with a fake GPU provider)
uv run ruff check .    # lint
uv run ruff format .   # format

Layout (src/gpu_vram_cleaner/):

License

Licensed under the Apache License 2.0.