Skip to content

gputaskman Shows Which Process Is Eating Your GPU Memory

A Python terminal tool that ties per-process GPU memory to specific models and KV-cache sizes has drawn 3,125 stars in 11 days.

Draftship Editorial4 min read

The gputaskman GPU task manager is trending on GitHub with 3,125 stars and 151 forks as of 2026-11-14, all gathered in the 11 days since the repository was created on 2026-11-03. It is MIT licensed, written in Python, and published as an installable terminal tool. The pitch in the project's GitHub repository is direct: Windows Task Manager says 15 GB of GPU memory is used, but not whether that is model weights, KV cache, a browser or the desktop compositor.

What the tool actually measures

The source describes gputaskman as an htop for what is inside the GPU rather than how full it is. It is a terminal UI that connects per-process GPU memory and engine activity with loaded runtimes, model files and LLM buffer breakdowns. Device totals, load, temperature, power and clocks appear where the platform exposes them.

The project is upfront about precision. Device and process counters are measurements, but classifications and the GGUF-derived breakdowns are estimates. A PyTorch or CUDA DLL alone cannot identify a model or prove training is running, and a mapped file does not prove a model is resident on the GPU. The stated scope is local GPU workloads, including decision models, vision, speech, LLM inference, training and fine-tuning, graphics and video. Cloud workloads are out of scope, and so are CPU-only and NPU-only jobs.

Install routes and platform coverage

The tool is on PyPI and installs with uv or pipx, with gtm as a shorter command. Python 3.11 or newer is required. For Windows machines without Python there are portable ZIPs for x64 and arm64, which are separate native builds rather than one universal executable; the whole archive has to be extracted so the _internal folder stays beside the binary. Both Windows builds are currently unsigned and no installer ships yet. Tagged releases also include Linux tarballs for x64 and arm64, which need glibc 2.35 or newer, and macOS on Apple Silicon is listed as supported.

Monitoring needs no CUDA Toolkit, administrator session, API key or model download. NVML comes with the NVIDIA driver and AMD ADL with supported AMD Windows drivers. On Windows the tool reads NVIDIA and AMD telemetry plus Intel WDDM activity. Native Linux covers NVIDIA, AMD on the amdgpu driver without ROCm or AMD SMI, Intel on i915, and Huawei Ascend 310P through npu-smi. Under WSL with interop on, per-process VRAM is matched from the Windows host and marked to show the match is inferred. --json and --once output make it usable from scripts, and gputaskman diff reports change in device usage over a run.

What the live checks showed

The source lists validation on an RTX 4080 SUPER and an RTX Spark N1X on Windows ARM64, plus an AMD Radeon 860M on Windows 11 Enterprise build 26300 with driver 32.0.22050.1003. On that AMD machine, per-process VRAM and GPU utilization were reported but temperature, memory activity and PCIe readings were unavailable, and Vulkan or ROCm workloads were not tested.

Two checks matter more for anyone running local models. In a Vast.ai Docker container on Ubuntu 22.04 with an RTX 3060 Ti and driver 580.82.09, three PyTorch jobs were matched to the correct container processes with their NVML sizes of 1.12, 2.52 and 1.59 GB, and an Ollama qwen3:0.6b runner was identified with its weights, KV-cache and compute split shown. On bare metal with CachyOS on kernel 7.3.3 and an RTX 4050 Laptop GPU, a llama-server running Qwen2.5-1.5B Q4_K_M at -ngl 99 -c 4096 showed 1.19 GB against 1218 MiB from nvidia-smi, and the exact CUDA0 buffers (935 MB of weights, 112 MB of KV cache) matched the tool's estimates. GPU utilization read 96% against 95% from nvidia-smi pmon.

That is the practical use: if you size a context window during local inference, the KV-cache line tells you what the extra context costs before you deploy.

Limits and open questions

The project calls itself early and experimental. Windows builds are unsigned, there is no installer, and code signing is not part of this initial distribution. Intel Arc B580 behaviour is covered only by mocked counters and has not been validated on Intel hardware. On native Linux the Spill and Paged columns stay blank because there is no equivalent commit counter, and JSON returns null rather than a zero. Container matching pairs NVML host PIDs with container processes by creation order, which the source marks with * or ? to signal how confident the match is. Huawei Ascend support came from a contributor's test on an Atlas 300I Duo with a vLLM workload and has not been independently validated, which sits oddly beside the statement that NPU-only workloads are out of scope.

What this means for AI app builders

The tool is for machines you own or rent, not for hosted apps. If you are shipping an app whose AI runs on someone else's infrastructure, gputaskman will not tell you anything about it, and a local reading of your laptop's VRAM says nothing about production. If you do want to add model calls to a hosted app, Draftship's guide to adding AI to your app covers that side of the work. Draftship is not affiliated with gputaskman.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.

Developer tools

GitHub Copilot Evals enters public preview

GitHub Copilot Evals is in public preview. It lets you create repeatable tests for skills, custom agents and MCP tools, then compare runs after a change.