GPU Taskman Shows Where Local LLMs Spend Your VRAM

If you run local AI models, you have probably stared at a task manager that says "15 GB used" and wondered what exactly is eating your graphics memory. GPU Taskman is a new open-source answer. Built by Microsoft VP Scott Hanselman, this terminal tool shows which processes are using your GPU, how much VRAM each one holds, and, for LLM workloads, how that memory splits between model weights and the KV cache.

• GPU Taskman is free, open-source software released under the permissive MIT license
• It runs on Windows with NVIDIA and AMD GPUs, on Linux and WSL, and on macOS with Apple Silicon
• Tom's Hardware tested the v0.3.0 portable Windows build, which needs no Python install
• For runners like llama.cpp and Ollama, it can break memory into weights, KV cache and compute buffers

What Does GPU Taskman Actually Show?

The project describes itself as an htop for what is inside your GPU, not just how full it is. Its terminal interface connects per-process GPU memory and engine activity with the runtimes and model files each process has loaded. Tom's Hardware highlights a screenshot where selecting llama-server reveals a Qwen3-0.6B model, with its weights, KV cache and parameter count broken out in the detail pane.

That matters because LLM memory use is famously hard to reason about. Model weights are only part of the bill: the KV cache grows with context length, and runtimes reserve extra buffers for compute. Seeing those pieces side by side makes it far easier to decide whether to drop to a smaller quantization, shorten your context window, or simply close the browser tab that is quietly holding a gigabyte.

The tool also surfaces something many users never notice: GPU-accessible memory living in system RAM. The README calls this out directly, with a Spill column that sorts processes by how much of their GPU allocation has landed in shared memory, a common cause of sudden slowdowns.

How Accurate Is GPU Taskman?

This is where the project earns trust. The README is unusually clear about what is measured and what is estimated. Device and process counters come straight from driver interfaces such as NVIDIA's NVML and AMD's ADL. Model identification and GGUF-derived memory breakdowns, however, are labeled as estimates that depend on supporting evidence. It even reports how each row was inferred in its JSON output.

The scope is deliberately local: graphics, video, speech, vision, LLM inference and fine-tuning workloads are all in, while CPU-only, NPU-only and cloud workloads are out. Hanselman calls it early, experimental software, and on social media he invited pull requests from the community.

How Do You Install GPU Taskman?

Windows users can grab a portable ZIP from the project's releases page, with separate x64 and ARM64 builds, then run it from a terminal. Anyone with Python 3.11 or newer can install it from PyPI using uv or pipx, and portable Linux builds are also published. Beyond the live interface, it offers one-shot text and JSON output, which makes it handy for scripts that log memory use during long inference runs.

If you are choosing a local inference stack, pair this with our guide to Ollama, vLLM and llama.cpp servers, and see how quantization can shrink memory needs in our Gemma 4 QAT VRAM breakdown. More tools live in our AI coverage.

Sources: GPU Taskman on GitHub — accessed November 8, 2026; Tom's Hardware — November 8, 2026.