The operating system for running AI on one GPU.
Chat models and ComfyUI share one NVIDIA card under one GPU manager. Models switch in about a second. Updates that fail roll themselves back.
Switch models in a second, not a coffee break.
elsieOS keeps models it is not using parked in pinned memory and copies them back to the card at the bus's full speed, instead of reading them from disk again.
Model switch
per request, answer included. Ollama on the same machine: 1.6-11 s.
Chat → ComfyUI
from a new image job arriving to the job starting on the card (0.27-0.86 s both ways).
Then it rolls back
a new version gets three starts to prove itself; after that the old one comes back by itself.
Ollama-compatible
apps made for Ollama talk to elsieOS without changes.
Time to switch to another chat model
A 24B and a 27B model that do not fit on the card together. Shorter is better.
One card. Every AI app. No fighting over memory.
A single-purpose system built on the Linux kernel and Debian, with nothing on it that is not there to run AI.
One referee for the graphics card
Chat models and ComfyUI take turns. A new image job parks idle chat models; a chat request waits for the job instead of crashing it.
gpu: shared
Models park, they do not unload
Weights move to pinned memory and back at ~50 GB/s, and the prompt cache survives, so the next answer starts right away.
park 0.29 s · wake 0.32 s
Updates that undo themselves
Two copies of the system. A new version has to start cleanly; if it fails three times the machine goes back to the old one on its own.
tries: 3 → rollback
Works with the apps you have
The same API as Ollama on the same port, plus the OpenAI-style one, so chat front ends, scripts and tools keep working.
POST /api/chat
Where elsieOS is now
Phase 4 of 9: the GPU manager and ComfyUI. Not installable yet; follow along for test builds.
News, test builds and benchmarks
Questions, ideas and results from your own card are welcome.