gpu-tpu-pedia

This skill should be used when users need to install, configure, or troubleshoot DeepEP (DeepSeek Expert Parallelism) on NVIDIA GPU systems. It covers the complete installation workflow including CUDA, DOCA-OFED, NVSHMEM with IBGDA support, and DeepEP itself. The skill is particularly useful for B200/H100/A100 GPUs with RoCE/InfiniBand networking, and includes comprehensive debugging capabilities for common installation failures.

创建和管理 Agent Teams 并行任务团队。当用户说"建个团队"、"拉个队伍"、"并行跑"、"create team"、"spawn agents"、"组队"时触发。

查看 Agent Teams 状态。当用户说"teams"、"团队状态"、"看看团队"、"team status"时触发。

Generate videos using Google Veo 3.1 on Vertex AI. Use when the user says "生成视频", "做个视频", "generate video", "帮我生成一段视频", "create a video", "视频生成", "text to video", "文生视频", or when you need to create video content.

Generate images using Google Imagen 4 on Vertex AI. Use when the user says "生成图片", "画一张", "generate image", "帮我画", "生成一张图", "create an image", "图片生成", or when you need to create visual content to explain concepts.

Deploy a Discord Bot that connects to Claude Code via persistent processes using Unix socketpair + stream-json (same mechanism as VSCode extension). Each user gets their own long-running Claude process with full interactive mode support (auto memory, CLAUDE.md, skills). Includes Whisper voice transcription, user whitelist, session history with dropdown switcher, and slash commands. Use when the user says "帮我建一个 Discord Bot", "setup discord bot", "搭建 Discord 机器人", "discord bot 设置", or "部署 discord bot".

安装配置 nginx 静态文件服务器，用于展示 Claude Code 生成的富内容（表格、图表、报告等）。当用户说"装nginx"、"配web服务"、"setup web server"时触发。

Discord 消息格式化规则。当 Claude Code 通过 Discord Bot 运行时自动适配 Discord 平台的消息风格，避免渲染问题。在 Discord 会话中始终遵循这些规则。当用户说"discord风格"、"discord格式"、"适配discord"时触发。

通过 tmux 远程控制交互式进程，编排多个并行任务。适用于在 GPU/TPU 实例上管理训练、推理、监控等长时间运行的进程。当用户说"tmux编排"、"并行跑"、"后台启动"、"tmux管理"、"多进程管理"、"帮我在tmux里跑"时触发。与 tmux-installer（安装配置）不同，本 skill 专注于运行时编排。

通过 Discord Bot 将任务报告发送到 Discord 频道。当用户说"discord通知我"、"ds通知我"、"发discord"、"发ds"等关键词时触发此技能。

This skill should be used when users need to install, configure, and customize tmux with Oh My Tmux (gpakosz/.tmux) on Linux systems. It covers tmux installation, Oh My Tmux setup, Tokyo Night color theme, Powerline separators, custom keybindings (split panes, window navigation, synchronize-panes), mouse mode, and clipboard integration. The skill replicates the standard team tmux configuration.

This skill should be used when users need to install, configure, and customize Zsh with Oh My Zsh on Linux systems. It covers Zsh installation, Oh My Zsh setup, theme configuration (agnoster), plugin management, and environment variable setup. The skill replicates the standard team shell configuration for consistent development environments across machines.

This skill should be used when users need to detect, format, and mount Google Cloud Local SSDs (NVMe) as a RAID0 array on /lssd. It covers SSD detection, RAID0 creation with mdadm, filesystem formatting, mounting, and HuggingFace cache directory setup. The skill is particularly useful for high-speed storage for ML model caching, checkpoints, and temporary data on Google Cloud VMs.

This skill should be used when users need to install, configure, debug, or run SGLang inference server on NVIDIA GPUs (especially B200/H100/A100). It covers installation from source, dependency management, environment setup, common error diagnosis and fixes, tensor parallelism configuration, and server startup/testing.

TPU v7 (Ironwood) 模型训练自动化。当用户说"帮我训练"、"跑一下训练"、"测试训练"等并指定 tpu-recipes 下的模型路径时触发。自动生成脚本、提交训练、收集结果并写文档。

将任务报告发送到飞书。当用户说"把报告发到飞书"、"飞书发给我"、"发飞书"等关键词时触发此技能。

管理客户贡献报告。用于记录对客户的技术支持、市场活动、培训活动等贡献，并维护中英文双版本报告。

通过 Server酱 将任务报告发送到微信。当用户说"微信发给我"、"发微信"、"发到微信"、"微信通知"等关键词时触发此技能。

Parallel SSH operations across multiple hosts using google_compute_engine key. Use this skill when the user provides a list of host IPs and wants to execute commands in parallel across them, such as starting prefill/decode nodes for distributed inference.

This skill should be used when users need to install, configure, debug, or run vLLM inference server on NVIDIA GPUs (especially B200/H100/A100). It covers installation from PyPI or source, dependency management, environment setup, common error diagnosis and fixes, tensor parallelism configuration, and server startup/testing. The skill automatically checks for LSSD mount status and DeepEP installation for MoE models.