docker-compose.gpu.yml

# GPU overlay:  docker compose -f docker-compose.yml -f docker-compose.gpu.yml up --build
#
# (12-factor V) TORCH_INDEX_URL selects CPU vs CUDA wheels at BUILD time, so
# reserving a GPU at run time is not enough on its own. This overlay previously
# added only the device reservation, which meant the command above ran the
# CPU-wheel image on GPU hardware unless someone had separately rebuilt with the
# CUDA index — silently, at the ~100-400x slowdown classify_TEXT.py warns about.
#
# The image name is also distinct. The CPU and GPU variants are different images
# built from the same Dockerfile with different args; sharing one name meant
# whichever was built last won, and `docker compose pull` could quietly replace
# one with the other.
#
# It is a LOCAL name (atrium-project#69, roadmap B2): no workflow publishes a GPU
# image, so the old `ghcr.io/…:<version>-gpu` could never be pulled, and
# `pull_policy: build` keeps `docker compose pull` from asking a registry for it.
# ATRIUM_RUNNER_IMAGE follows the name in both places: merged onto
# docker-compose.yml, this service otherwise kept the CPU image's name, so a GPU
# run's paradata recorded an image it did not run (roadmap B3's compose half).
services:
  alto:
    image: atrium-ocr-postprocess-gpu:local
    pull_policy: build
    build:
      args:
        TORCH_INDEX_URL: ${TORCH_INDEX_URL:-https://download.pytorch.org/whl/cu124}
        ATRIUM_RUNNER_IMAGE: atrium-ocr-postprocess-gpu:local
    environment:
      NVIDIA_VISIBLE_DEVICES: all
      ATRIUM_RUNNER_IMAGE: atrium-ocr-postprocess-gpu:local
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]