version: "1.0.0" name: gpu-mode description: Coordinate which vLLM owns the single DGX GPU before using a local vLLM endpoint. Checks the active mode (prod/ollama/code/research/idle/BROKEN-BOTH) read-only, and switches only on explicit request. Use before pointing any tool at a local vLLM (http://<dgx>:9000/v1 or :8003/v1), when a vLLM call fails to connect, or when asked to bring the coder-next or autoresearch vLLM up or down. Switching is shared-state — never switch without explicit approval.

gpu-mode

The DGX has one GB10 GPU; the coder-next and autoresearch vLLMs can't both hold it (both want ~90% of VRAM). gpu-mode-swap.sh is the authoritative coordinator. This skill wraps it: check freely, switch only on explicit go.

Connect

On the DGX, ~/bin/gpu-mode-swap.sh is a wrapper that runs the real script as the ops account (sudo -u ops /home/ops/agentic-ai-homelab/infra/dgx/bin/gpu-mode-swap.sh), which owns the vLLM stacks and their .env files. Reach it over the tailnet with ssh (non-interactive, so always the absolute path; the gpu-mode zsh alias is not loaded in agent shells):

ssh dgx-llm-1 '~/bin/gpu-mode-swap.sh <args>'

Check the mode — read-only, always safe

Before ANY local vLLM call, confirm the right stack owns the GPU:

ssh dgx-llm-1 '~/bin/gpu-mode-swap.sh --mode-only'    # → prod | ollama | code | research | idle | BROKEN-BOTH
ssh dgx-llm-1 '~/bin/gpu-mode-swap.sh status --json'  # full machine-readable state

• prod — the production serving vLLM on :8003 (infra/vllm/prod-vllm), plus the

co-resident translation vLLM on :8005. Ollama and its ollama-metrics exporter are stopped. The DGX's normal state (2026-10-17).

• ollama — the podcast_scraper pipeline's Ollama mode: no vLLM, Ollama warm on :11434.
• code — coder-next up on :9000
• research — autoresearch up on :8003
• idle — both down; GPU free for ML training / Ollama
• BROKEN-BOTH — both listening (bad state); needs a human, do not pile on

If the mode already matches what you need, proceed. If not, see below — but do not switch on your own.

Switch the mode — SHARED-STATE, explicit approval required every time

Switching brings one vLLM up and the other down; it interrupts whatever is using the GPU right now (a sweep, a training job, Ollama). Treat the DGX as production:

• Never switch as a side effect of "I need the endpoint." Ask first.
• idle does not mean the GPU is idle — a training job or Ollama can hold it

directly while gpu-mode-swap reports idle. Confirm with the operator before taking the GPU.

• Only after an explicit go, and by absolute path:

``bash ssh dgx-llm-1 '~/bin/gpu-mode-swap.sh code --json' # coder-next up, autoresearch down ssh dgx-llm-1 '~/bin/gpu-mode-swap.sh research --json' # autoresearch up, coder-next down ssh dgx-llm-1 '~/bin/gpu-mode-swap.sh idle --json' # both down ``

Report

• State the current mode plainly. If a switch happened, show the --json result

and the resulting mode.

• On BROKEN-BOTH or a failed wait_for_port, stop and surface it — follow the

troubleshooting in docs/recipes/gpu-mode-swap.md; do not retry blindly.