Playbooks · Coding tasks
gputaskman: see what your LLM is holding
Written November 12, 2026
Your task manager says 15 GB of GPU memory is used. gputaskman answers the next question: used by what — model weights, KV cache, compute buffers, your browser, or the desktop itself — process by process.
What it shows
gputaskman, by Scott Hanselman, describes itself as an htop for what is inside your GPU, not just how full it is. It is a terminal tool that connects per-process GPU memory and engine activity with loaded runtimes, model files and LLM buffer breakdowns. For a local model, that breakdown can separate weights from KV cache and compute buffers, which is the difference between choosing a smaller quantization, cutting context length, or closing the other process that is holding the memory. It also covers graphics, video, training and fine-tuning workloads, and it can compare snapshots and record a session.
Where it runs, and how to get it
It is on PyPI, so no clone is needed: uv tool install gputaskman or pipx install gputaskman, then run gputaskman (a shorter gtm command is also installed). It needs Python 3.11 or later. Portable Windows downloads, separate x64 and ARM64 ZIP files, need no Python at all. The supported ground is Windows with NVIDIA and AMD telemetry and Intel activity through WDDM, Linux and WSL, and macOS on Apple Silicon, with the detail varying by platform: per-process memory in WSL, for example, is matched from the Windows host and marked as matched, while on native Linux it is measured directly. The project is MIT licensed. As read for this page, the repository shows about 2,599 stars and was created on November 3, 2026.
Read its numbers the way it asks you to
The README is unusually plain about its own limits, and the page follows it. gputaskman calls itself early, experimental software. Device and process counters are measurements, but classifications and GGUF-derived breakdowns are estimates. A mapped model file does not prove the model is resident on the GPU or being exercised, and a PyTorch or CUDA library alone cannot identify a model. On Linux, only processes you can inspect are listed; other users' processes need sudo. Unavailable counters are shown as unavailable, or as null in JSON output, rather than as a measured zero. Cloud workloads are out of scope: this is a tool for the GPU in front of you.
Limits
- The Windows portable builds are currently unsigned, and there is no installer. That is the project's own description of its initial distribution.
- Model identification needs supporting evidence, and some platform and vendor combinations rest on mocked counters rather than hardware the project has run on. The README names which is which, case by case.
- It reports memory and activity. It does not free memory, move processes, or tell you which model to run; it gives you the numbers to make that call yourself.
Related playbooks
- Mistral Vibe 3.1.0: see what fills the context
- Claude agents now dream: memory that consolidates
- Keep Claude Code memory under 200 lines
- Codex memory v2: import memories from other agents
Sources
Back to: Coding playbooks hub · Also: BotPsy 0.8B: an agent model that fits on-device