docker-compose.gpu.yml

# Oneirodex Compose overlay — hand the artwork sidecar an NVIDIA GPU.
#
# OPT-IN. Nothing in the base stack requests a GPU, and that is deliberate: the
# reference deployment is Unraid on a box with no discrete GPU, and a `deploy`
# reservation on a host without a working NVIDIA driver does not degrade — it
# hard-fails the container create with
#
#   error running prestart hook #0 ... nvidia-container-cli: initialization
#   error: nvml error: driver not loaded
#
# which fails the whole `compose up` and reads like the stack is broken when
# only the optional artwork sidecar is. So the reservation lives here, in a file
# you have to ask for, instead of in docker-compose.yml where every deployment
# inherits it.
#
# Use it when the DOCKER HOST ITSELF has the NVIDIA driver and the NVIDIA
# Container Toolkit. If the GPU is on a *different* machine, do not use this at
# all — run SD.Next / AUTOMATIC1111 / Forge over there and point
# AI_ARTWORK_URL at it. `oneirodex/utils/ai_artwork.py` only ever speaks HTTP to
# that URL, so a GPU box elsewhere on the LAN needs no compose changes here.
#
# Enable, on a Linux host (Unraid, NAS, plain Docker) — in .env:
#
#   COMPOSE_FILE=docker-compose.yml:docker-compose.gpu.yml
#
# On Windows the default separator is ";" rather than ":", so either use ";" or
# pin the separator explicitly:
#
#   COMPOSE_PATH_SEPARATOR=:
#   COMPOSE_FILE=docker-compose.yml:docker-compose.gpu.yml
#
# Verify the host can do it at all BEFORE enabling — this is the same check the
# prestart hook makes, and it fails in the same way:
#
#   docker run --rm --gpus all nvidia/cuda:12.4.0-base-ubuntu22.04 nvidia-smi
#
# The sidecar is still behind the `artwork` profile either way, so this overlay
# only takes effect on `docker compose --profile artwork up -d`.

services:
  sdnext:
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]