boxmoe_header_banner_img

Hello! 欢迎来到阿虪的开发世界~

加载中

文章导读

gputaskman:给 GPU 装个 htop,终于能看清显存被谁吃掉了


avatar
阿虪 16 11 月, 2026 102
gputaskman

中文版

gputaskman(GitHub:https://github.com/shanselman/gputaskman ,2026-11-03 创建,今晨抓取时已 3,342★)是一个“给 GPU 用的 htop”:终端界面,告诉你显存到底被什么吃掉了,而不是只报一个“已用 15GB”。

按 README 的说法,它支持 Windows(NVIDIA/AMD 遥测和 Intel WDDM)、Linux 和 WSL(全平台 NVIDIA,原生 Linux 上还有 AMD 和 Intel),以及 Apple Silicon 的 macOS。任务管理器只会说“用了 15GB”,它要回答的是:被谁用的?模型权重?KV 缓存?浏览器?桌面合成器?为什么这么多显存跑进了系统内存?它把每个进程的显存占用、引擎活动和已加载的运行时、模型文件、LLM 缓冲区明细连起来看——本地跑大模型、推理、训练微调(包括 Unsloth)、图形和视频都算在内。装法是 pip install(PyPI 有包)。MIT 开源,作者是微软的 Scott Hanselman。

需要注意:README 自己注明这是早期实验性软件,进程计数是实测值,模型识别和 GGUF 拆解是估算;它不转储显存、也不枚举游戏贴图;只管本地 GPU 工作负载,不管云服务和纯 CPU/NPU 任务。

为什么值得了解

  • 本地跑大模型的人都有这个痛点:显存一满就报错,却不知道是谁占的——它把“模型权重 / KV 缓存 / 计算缓冲”拆开给你看,定位问题快得多
  • 跨平台覆盖全:Windows、Linux、WSL、Apple Silicon Mac 一把抓,换机器不用换工具
  • 作者是 Scott Hanselman,微软资深布道者,项目质量和维护有保障
  • 11 月 3 日才创建,不到两周就 3,300+ 星:本地 LLM 玩家都在找这种工具

English Version

gputaskman (GitHub: https://github.com/shanselman/gputaskman , created 2026-11-03, 3,342 stars at this morning’s fetch) is an “htop for your GPU”: a terminal UI that shows you what is actually eating your VRAM, instead of just reporting “15 GB used”.

Per the README, it supports Windows (NVIDIA/AMD telemetry and Intel WDDM activity), Linux and WSL (NVIDIA everywhere, AMD and Intel on native Linux), and macOS on Apple Silicon. Task Manager says “15 GB used”; this tries to answer: used by what — model weights, KV cache, your browser, the desktop compositor, and why so much GPU memory ended up in system RAM? It connects per-process GPU memory and engine activity with loaded runtimes, model files and LLM buffer breakdowns — local LLM inference, training and fine-tuning (including Unsloth), graphics and video all belong here. Install it from PyPI. MIT licensed, by Microsoft’s Scott Hanselman.

Two notes: the README itself calls it early, experimental software — process counters are measurements, while model identification and GGUF breakdowns are estimates; it does not dump VRAM or enumerate game textures. And it covers local GPU workloads only, not cloud services or CPU/NPU-only tasks.

Why it’s worth knowing

  • Anyone running models locally knows this pain: VRAM fills up and errors out, and nobody can say who took it — splitting the usage into model weights, KV cache and compute buffers makes the culprit obvious much faster
  • Full cross-platform coverage: Windows, Linux, WSL and Apple Silicon Macs in one tool, so switching machines never means switching tools
  • By Scott Hanselman, a longtime Microsoft advocate — project quality and maintenance you can rely on
  • Created on November 3 and already past 3,300 stars in under two weeks: local LLM users have clearly been looking for exactly this


评论(0)

查看评论列表

暂无评论


发表评论

表情 颜文字

插入代码