GPU Taskman shows per-process VRAM usage in GB and helps track down memory pressure

You launch a local LLM, increase the context length—and run out of free GPU memory. Before switching to a smaller model or buying another graphics card, it helps to understand where the gigabytes went. GPU Taskman collects per-process memory usage and helps tie it to a specific workload: a model, a browser, or another task running alongside it.

For developers, the benefit is straightforward: a picture of memory usage before and after a settings change. If the problem appeared after increasing the context length, comparing snapshots can help establish which process started using more memory. An estimate of the model's memory breakdown can then suggest which hypothesis to test next.

What the task manager shows

GPU Taskman runs in the terminal. It brings together information on process GPU memory usage, GPU activity, loaded libraries, and model files. In supported configurations, it can show which stack an application uses: CUDA, Vulkan, or Direct3D. For local LLMs, it provides model identification and estimates of memory allocated to weights, the KV cache, and working buffers.

These details have different levels of reliability. The memory counter reports measured usage. Identifying the model and estimating the breakdown of that usage require additional clues. The presence of a CUDA library alone does not explain which model is loaded or what it is currently doing. The author also explicitly notes the software's experimental status in the project documentation.

It is therefore useful to read the screen in two passes. First, find the process with high memory usage. Then look at the explanations the utility has been able to gather for it. This approach helps avoid mistaking a plausible label for a direct measurement.

How to check where the memory is going

Consider a common situation: a model starts successfully, but after you change its settings, the application begins reporting out-of-memory errors. Save a baseline snapshot, change one parameter, and capture the readings again. GPU Taskman provides the --once option for a single readout, --json for machine-readable output, and diff for comparisons.

Context length can be the first parameter to test. The KV cache stores data the model uses when processing a sequence, and its size is related to the context. However, the exact behavior depends on the engine and settings: memory may be reserved in advance. It is therefore useful to compare the same stages of application execution, such as the state after loading the model or after running the same prompt.

If you change the model, context size, and other settings at the same time, the increase in usage will be harder to explain. The practical purpose of this comparison is to narrow down the cause. It can also help reveal another process that continues holding memory after your task has finished. The monitor itself does not free memory or fix model settings.

Where the limitations begin

WSL requires particular attention. Memory information comes from Windows, and the association between individual readings and Linux processes has to be reconstructed. The program labels confident matches and guesses, and may display n/a when there is insufficient data. Check that label before drawing conclusions about a particular process.

The project lists support for Windows, Linux, WSL, and Apple Silicon, but the availability of individual metrics depends on the platform, driver, and graphics card. A shared interface does not mean the same depth of visibility on every machine. The LLM memory breakdown should also be treated as a diagnostic clue: the utility does not read the contents of VRAM or provide a complete inventory of the objects stored there.

Install with uv tool install gputaskman or pipx install gputaskman, then launch gputaskman or gtm. A good starting point is one reproducible test using your own workload. If the snapshots help explain memory usage, the tool is already useful: they provide a basis for changing a specific setting and checking the result before spending time and money on replacement hardware.