Quickstart

Four commands. The whole session is: claim the GPU, tunnel, work, release.

1. Claim the GPU

ssh f-nui-zeph-01 comfy-up

This stops the two resident services, waits for the driver to actually release the memory, starts ComfyUI, and prints how much VRAM you ended up with. Expect output like:

19:05:13  stopping scholar-api.service (replica query API, ~3.4 GB VRAM)
19:05:14  unloading qwen3:4b from ollama (~3.9 GB VRAM)
19:05:15  VRAM in use: 7359 MiB -> 2 MiB
19:05:22  ComfyUI up after 8s (pid 384214), 15.3 GiB VRAM free

It is safe to run twice — if ComfyUI is already up it says so and exits without touching anything.

2. Tunnel from your own machine

ComfyUI binds loopback only. It is not on the network, and it has no login of its own, so the tunnel is the access control.

ssh -N -L 8188:127.0.0.1:8188 f-nui-zeph-01

Leave that running and open http://127.0.0.1:8188 (opens in new tab) in your browser.

3. Generate

Load a workflow from Workflows and hit Run. First generation of a session is slower — the model has to be read from disk into VRAM.

4. Hand the card back

ssh f-nui-zeph-01 comfy-down

This stops ComfyUI, restarts the paper-RAG API, re-pins the LLM, and then verifies the API is answering before it reports success. Do not skip it — the RAG stays dark until you do.

If you forget step 4

Nothing breaks and nothing is lost. Replication from the primary keeps running the whole time (see Sharing the GPU ), so the corpus stays current — you just cannot query the replica until someone runs comfy-down.