GPU Task Manager
An htop for what's inside your GPU — per-process VRAM, and which model is eating it.
Project intro
Task Manager tells you 15 GB of GPU memory is used. GPU Task Manager tries to answer the next question: used by what? It is a terminal UI that ties per-process GPU memory and engine activity to loaded runtimes, mapped model files, and LLM buffer breakdowns — model weights, KV cache, and compute buffers, process by process. It covers Windows with NVIDIA, AMD, and Intel graphics, Linux and WSL, and macOS on Apple Silicon.
Install it from PyPI with one command, or grab a portable Windows build that needs no Python at all, and it starts naming what it finds: llama.cpp, LM Studio, Ollama, and stable-diffusion.cpp workloads, with the model, architecture, quantization, context size, and file behind each one. The README walks through a real example on an RTX 4080 SUPER, where a Qwen3-0.6B server holds 604 MB of weights and 448 MB of KV cache.
More than 3,400 developers starred it in its first thirteen days, version 0.4.3 is the latest release, and it was covered by Neowin and BetaNews within a week of launch. Scott Hanselman built it to find out where his own GPU memory was going, and it is MIT licensed with no telemetry.
It is early, experimental software, and the README says so plainly: device and process counters are measurements, while model identification and GGUF-derived breakdowns are estimates, not ground truth. It does not dump VRAM or enumerate a game's textures, and the Windows builds are unsigned with no installer yet.
Actions
Tech stack
Project info
- Project category
- Developer Tools
- Project status
- Online
- Collaboration intent
- None