Infrastructure · since 2026live

Doum

One GPU, many AI jobs, taking turns.

Our open-source GPU queue. The lab's NVIDIA GB10 runs local LLMs, image and video models and voice training; Doum makes them take turns instead of crashing each other.

Local AI labsShared GPU workstationsAI coding agents

Visit github.com/MoeEldouma/doum →

37 sfrom idle to a 120B model's first answer, loaded on demand

1,000,629simulated events, zero deadlocks

Apache-2.0open source, v0.1.0

How it works

From input to finished result.

A job asksAn LLM, a render, a training run or an AI coding session

Reserve memoryEach job books GPU memory shards and waits its turn

PrioritiseUrgent work pauses long jobs, which resume on their own

Run safelySafety checks before every job: memory, heat, cool-down

Watch itLive web panel, terminal UI, metrics and phone alerts

What it does

The parts that make Doum useful.

Fair queue

Jobs reserve memory and wait in order; urgent work pauses long jobs, which pick up where they left off.

Memory shards · pause and resume

Talks to every app

A ComfyUI plugin, local LLMs loaded on demand, and MCP tools plus a hook so AI coding sessions queue too.

ComfyUI · llama-swap · MCP

See everything

A live web panel with a timeline of parallel runs, a terminal UI, Prometheus metrics and phone alerts.

Web panel · TUI · Prometheus · ntfy

Safe by default

Local-only API with a host check and owner-only state; background jobs survive restarts.

Local-only · survives restarts

Models and tools

What Doum runs on.

NVIDIA GB10

ComfyUI

Rust

Prometheus

Names and logos belong to their owners and identify the models and tools we use; no endorsement is implied.

Built so far

Engineering that is already in place.

Runs the lab's own GPU every day
Written in Rust with an AI pair programmer
Open source on GitHub

Where it's going

The next steps, in order.

Freeze in placePause GPU jobs exactly where they are with cuda-checkpoint
ClustersA Slurm backend for more than one machine