Skip to content

downstairs-wsl-gpu.md

Latest commit

 

History

History
55 lines (38 loc) · 1.76 KB

downstairs-wsl-gpu.md

File metadata and controls

55 lines (38 loc) · 1.76 KB
Raw

Example: GPU inference host in WSL (public-safe)

A common home-lab split is:

  • Workstation — IDE + local-coding-slm MCP
  • Second PC — Windows, Ubuntu in WSL2, NVIDIA GPU, Ollama

This file uses placeholders only. Put the real SSH name, port, and LAN address in a gitignored .env or *.local.md. Never commit them.

Why SSH instead of a LAN bind

Keep Ollama on WSL localhost. From the workstation, forward the API:

ssh -N -T -o ExitOnForwardFailure=yes -p <ssh-port> \
  -L 127.0.0.1:11436:127.0.0.1:11434 user@<wsl-host>

Then on the workstation (gitignored .env):

OLLAMA_BASE_URL=http://127.0.0.1:11436

11436 is an example unused port on the workstation. 11434 is Ollama's loopback port inside WSL. <wsl-host> is a name or address you already use. Do not scan for it. <ssh-port> is often not 22 when Windows already owns port 22 — use whatever ssh you already use to reach that WSL.

Documentation stand-ins (not a real household):

ssh -N -T -o ExitOnForwardFailure=yes -p 2222 \
  -L 127.0.0.1:11436:127.0.0.1:11434 user@wsl-gpu.example

wsl-gpu.example and 192.0.2.0/24 are examples. They are not this lab.

On the WSL host (when it is on)

  1. Install a current Ollama. Distro packages can be too old.
  2. Pull the starter tags and confirm ollama ps after one short chat.
  3. Leave Ollama on 127.0.0.1:11434 if you use the SSH forward above.
  4. The PC may be powered off when unused. Treat "no route" as host down before changing addresses.

Do not port-forward 11434 or 22/2222 through a public router.

What not to commit

  • Real hostnames, Windows computer names, usernames, or RFC1918 addresses
  • A working OLLAMA_BASE_URL that identifies the house
  • Screenshots of ssh or ip addr from that machine