fal-ai-community/smoltop

smoltop is a terminal dashboard for monitoring NVIDIA GPUs. If you have a machine with one or more NVIDIA graphics cards, say a server running machine learning jobs, smoltop shows you a live, updating view of what those cards are doing inside your terminal window. For each GPU, smoltop displays a range of live metrics: how hard the processor is working, memory usage, power draw, temperature, fan speed, and data transfer rates. It shows small history graphs for utilization and memory so you can see trends over time. If a card starts to slow down due to heat or power limits, a red warning badge appears immediately. The process table lists every program using a GPU. If a single job spans multiple cards, it appears as one row with the total memory combined. The table also shows the user, CPU usage, process age, and whether the process is running inside a Docker or Kubernetes container. A status line at the bottom tallies memory usage by user, so you can see who is consuming the most resources. The dashboard adjusts its layout automatically based on your terminal size, packing more detail into larger windows and condensing into fewer lines when space is tight. You can cycle through display densities manually by pressing a key. It includes matching dark and light themes, and can detect your terminal's background color by default, or you can pick a theme explicitly. You can control processes directly from the table. Pressing a key sends a termination signal to stop a job, with a confirmation popup showing exactly what will be stopped. Another key suspends and resumes a process instantly, which is useful for pausing work without canceling it. You can also filter the process list, sort by various columns, and open a detail view for any process to see its full command line and per-GPU memory breakdown. The tool runs on Linux with the NVIDIA driver loaded and requires Python 3.10 or newer. It is not yet on PyPI, so you install it directly from GitHub using a tool called uv, pipx, or pip. A demo mode simulates an eight-GPU system, which is handy if you want to try it without any actual hardware present.