Menu
Kill GPU-occupying Python/vLLM processes to free GPU memory. Use when nvidia-smi shows memory in use by dead or unwanted processes.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| name | kill-gpu |
| description | Kill GPU-occupying Python/vLLM processes to free GPU memory. Use when nvidia-smi shows memory in use by dead or unwanted processes. |
nvidia-smi --query-compute-apps=pid,used_memory,name --format=csv,noheader
pgrep -a python; pgrep -a vllm
kill -9 <PID>
nvidia-smi | grep -E "MiB.*MiB"
nvidia-smi --query-compute-apps shows a PID that no longer exists in pgrep, it's a zombie CUDA context — killing the parent process or waiting a few seconds usually clears it.Set up a model training/inference environment from scratch (uv venv + inference engine + training smoke test). Use when preparing to train or serve a model on a new machine/project, or when python/torch/vllm imports fail. Tool-agnostic flow; uv-first.
Generate a slide-by-slide Markdown PPT script (演讲文稿) from an existing document (PDF thesis, LaTeX source, research paper, project report, or meeting notes). The script maps one-to-one to PPT pages and captures: slide copy, figure references, and speaker notes. Use when the user asks to: - 写 PPT 文稿 / 起 PPT 稿 / 做 PPT 设计文稿 - Create a presentation script for thesis defense, research progress report, or project review - Convert an academic paper, thesis, or report into a structured PPT script - Draft speaker notes or narration for a presentation Do NOT use for actually implementing the .pptx file — use the pptx skill for that.