Agent Skills › NVIDIA/dgx-spark-playbooks › dgx-station-mig

dgx-station-mig

GitHub

用于检查 NVIDIA DGX Station GB300 的 MIG 状态、驱动配置，生成布局计划并执行变更。涵盖启用、禁用、分区及故障排查场景，强调安全验证与显式确认。

nvidia/playbook-dgx-station-ai-skills/assets/skills/dgx-station-mig/SKILL.md NVIDIA/dgx-spark-playbooks

触发场景

用户请求启用、禁用或重新配置 MIG 实例需要获取 MIG UUID 或进行 MIG 相关故障排查

安装

npx skills add NVIDIA/dgx-spark-playbooks --skill dgx-station-mig -g -y

更多选项

非标准路径

npx skills add https://github.com/NVIDIA/dgx-spark-playbooks/tree/main/nvidia/playbook-dgx-station-ai-skills/assets/skills/dgx-station-mig -g -y

不安装直接使用

npx skills use NVIDIA/dgx-spark-playbooks@dgx-station-mig

指定 Agent (Claude Code)

npx skills add NVIDIA/dgx-spark-playbooks --skill dgx-station-mig -a claude-code -g -y

安装 repo 全部 skill

npx skills add NVIDIA/dgx-spark-playbooks --all -g -y

预览 repo 内 skill

npx skills add NVIDIA/dgx-spark-playbooks --list

SKILL.md

Frontmatter

{
    "name": "dgx-station-mig",
    "description": "Inspect NVIDIA DGX Station GB300 MIG state and installed-driver profiles, create a digest-bound layout plan, disclose disruption and restoration, and apply an approved still-valid plan. Use when the user asks to enable, disable, partition, reconfigure, inspect, or troubleshoot MIG instances or needs MIG UUIDs. Never assume static profile IDs or terminate GPU clients."
}

DGX Station MIG

Treat inspection and planning as read-only. Treat every apply as disruptive.

Workflow

Run scripts/dgx-assist system inspect --json.
Search the playbooks for the user's MIG concern and cite relevant results.
Read compatibility.capabilities.mig_inspection. If it is false, stop and explain that the detected release has no MIG inspection profile.
Run mig inspect and mig profiles. Use only profiles reported by the installed driver. This read-only step is available on the recognized Software 1.0 and Software 2.0 profiles.
Before planning, require compatibility.capabilities.mig_mutation to be true. If it is false, report the observed state but do not suggest a layout or work around the restriction.
Require the desired exact layout. Do not choose a layout from model names or nominal memory sums.
Run mig plan --layout "<layout>".
Present current and proposed state, exact commands, active clients, privilege, workload impact, reset/reboot possibility, and restoration commands.
If any client is active, stop. Tell the user dgx-assist will not terminate it.
Obtain explicit confirmation immediately before mutation.
Run mig apply --plan-id "<id>" --yes.
Report every verified MIG UUID. On partial failure, stop and present only the recorded restoration plan.

Safety requirements

Never use a hardcoded GPU index, profile table, or assumed placement.
Never destroy instances, reset a GPU, reboot, or change MIG mode outside an approved plan.
Never kill a GPU process or take over a service.
Invalidate the plan when clients, mode, profiles, release, driver, or instances change.
Do not improvise further mutations after partial failure.
Do not treat Fabric Manager as a normal Station prerequisite.
Treat --yes only as approval already obtained.

Read references/workflow.md before planning or applying a layout.