Rycen7822/nvclean

Latest recorded source activity: May 6, 2026 at 19:05

SkillsMP catalog refreshed: November 9, 2026 at 14:22

skills collected: 2

GitHub stars: 1
GitHub forks: 0

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/Rycen7822/nvclean.

Repository README

Repository README source: README.md

Read the source README

Expand full README

# nvclean `nvclean` is a compact `nvidia-smi`-style CLI written in C. It queries NVIDIA GPUs through NVML directly and does not spawn or parse `nvidia-smi`. The output is intentionally short: time, driver/NVML/CUDA driver versions, per-GPU status, and active compute/graphics processes. ## Agent-friendly output `nvclean` and `nvmini` are designed for coding agents and terminal assistants that repeatedly inspect GPU state. Their compact, line-oriented output avoids the wide ASCII table produced by `nvidia-smi`, which can waste context tokens when copied into an LLM prompt. In one GPT-5.x & O1/O3 tokenizer check on May 7, 2026: | Tool | Tokens | Characters | Token reduction vs `nvidia-smi` | Character reduction vs `nvidia-smi` | | --- | ---: | ---: | ---: | ---: | | `nvidia-smi` | 317 | 1773 | baseline | baseline | | `nvclean` | 210 | 502 | 34% | 72% | | `nvmini` | 41 | 72 | 87% | 96% | Screenshots from that tokenizer check: <p> <img src="assets/tokenizer/nvidia-smi.png" alt="nvidia-smi tokenizer usage" width="32%"> <img src="assets/tokenizer/nvclean.png" alt="nvclean tokenizer usage" width="32%"> <img src="assets/tokenizer/nvmini.png" alt="nvmini tokenizer usage" width="32%"> </p> For agent workflows, prefer: ```bash nvclean ``` Fall back to `nvidia-smi` only when `nvclean` is unavailable or when you need fields outside this compact status view. For even tighter polling output, use `nvmini`. It omits time, driver details, PCI/display/ECC/compute-mode fields, GPU names, and process memory. The fixed field order is: ```text g<gpu> <used>/<total>MiB <util> <temp> <power> p<gpu> <type> <pid> <process> ``` Example: ```text g0 10084/32607MiB 90% 66C 573.5W p0 C 645937 python3.12 p0 G 28 Xwayland ``` ## Requirements - NVIDIA GPU with a working NVIDIA driver - NVML runtime library - NVML header file - C compiler and `make` On WSL, the common paths are: ```bash /usr/local/cuda/include/nvml.h /usr/lib/wsl/lib/libnvidia-ml.so.1 ``` If `nvml.h` is missing on WSL, install the CUDA Toolkit package, not the Linux display driver. ## Build ```bash make ``` This builds both `nvclean` and `nvmini`. For WSL, the default Makefile settings use: ```bash NVML_INCLUDE=/usr/local/cuda/include NVML_LIBDIR=/usr/lib/wsl/lib ``` For a regular Linux install where NVML is under `/usr/lib/x86_64-linux-gnu`, build with: ```bash make NVML_LIBDIR=/usr/lib/x86_64-linux-gnu ``` If your system only has `libnvidia-ml.so.1` and not the linker symlink `libnvidia-ml.so`, build with the library path directly: ```bash make LDLIBS=/usr/lib/wsl/lib/libnvidia-ml.so.1 ``` ## Install ```bash make install ``` This installs to `~/.local/bin/nvclean` by default. Override `PREFIX` if needed: ```bash make install PREFIX=/usr/local ``` `make install` installs both `nvclean` and `nvmini`. ## Codex skills This repository includes two small Codex skills: - `skills/nvclean/SKILL.md`: prefer `nvclean` over `nvidia-smi` - `skills/nvmini/SKILL.md`: prefer `nvmini` for the shortest routine GPU checks Prompt for an agent to install only `nvclean` and the `nvclean` skill: ```text Install nvclean only. Use https://github.com/Rycen7822/nvclean. Clone or update it under /tmp/nvclean, build with make nvclean, install ./nvclean to ~/.local/bin/nvclean, ensure ~/.local/bin is on PATH for future shells, copy skills/nvclean/SKILL.md to ~/.codex/skills/nvclean/SKILL.md, then verify with command -v nvclean and nvclean. Do not install nvmini or the nvmini skill. On WSL, do not install Linux NVIDIA display drivers. If linking fails because libnvidia-ml.so is missing, retry with make nvclean LDLIBS=/usr/lib/wsl/lib/libnvidia-ml.so.1. ``` Prompt for an agent to install only `nvmini` and the `nvmini` skill: ```text Install nvmini only. Use https://github.com/Rycen7822/nvclean. Clone or update it under /tmp/nvclean, build with make nvmini, install ./nvmini to ~/.local/bin/nvmini, ensure ~/.local/bin is on PATH for future shells, copy skills/nvmini/SKILL.md to ~/.codex/skills/nvmini/SKILL.md, then verify with command -v nvmini and nvmini. Do not install nvclean or the nvclean skill. On WSL, do not install Linux NVIDIA display drivers. If linking fails because libnvidia-ml.so is missing, retry with make nvmini LDLIBS=/usr/lib/wsl/lib/libnvidia-ml.so.1. ``` ## Examples `nvclean`: ```text Time: Thu May 7 00:28:37 2026 NVML: 595.71.01 | Driver: 596.36 | CUDA driver: 13.2 GPU 0: NVIDIA GeForce RTX 5090 Bus ID: 00000000:02:00.0 Persistence: On | Display: On | ECC: N/A Fan: 37% | Temp: 47C | Perf: P0 Power: 101W/600W | Memory: 6962MiB/32607MiB | Utilization: 7% Compute mode: Default Processes: GPU 0 | PID 28 | Type G | /Xwayland | Memory N/A GPU 0 | PID 629519 | Type C | /python3.12 | Memory N/A ``` Process memory may be reported as `N/A` on WSL/Windows WDDM systems because that NVML field is not always available there. `nvmini`: ```text g0 10084/32607MiB 90% 66C 573.5W p0 C 645937 python3.12 p0 G 28 Xwayland ``` ## License MIT

Loading the repository README…

Skills in this repository

Network & Computer Systems Administrators

1 occupation category · 100% classified

skill

occupation

description

updated

nvmini

Network & Computer Systems Administrators

run nvmini for shortest NVIDIA GPU checks; fall back if needed.

2026-05-06

nvclean

Network & Computer Systems Administrators

run nvclean instead of nvidia-smi; fall back if needed.

2026-05-06

Showing 2 of 2 collected skills.