docker-compose.gpu.yml

Rama main · réplica de lectura del repositorio original.

# GPU for the encoder. An overlay rather than part of the base file, because a
# device reservation on a machine with no GPU is a hard failure at `up`, and
# the default has to be a stack that starts anywhere:
#
#   docker compose -f docker-compose.yml -f docker-compose.gpu.yml up -d --build
#
# Or set it once, and every plain `docker compose` command picks it up:
#
#   COMPOSE_FILE=docker-compose.yml:docker-compose.gpu.yml
#
# Needs the NVIDIA Container Toolkit on the host. Check with:
#
#   docker run --rm --gpus all nvidia/cuda:12.8.0-base-ubuntu24.04 nvidia-smi
#
# The encoder falls back to CPU on its own if the GPU is not visible, so a
# misconfiguration here is slow rather than broken — check the startup line,
# which names the device it loaded on.
services:
  embedder:
    build:
      args:
        # Must match the host driver. `nvidia-smi` prints the CUDA version it
        # supports in the top right; cu130 wants 580 or newer.
        TORCH_INDEX: ${FOILSTACK_TORCH_INDEX:-https://download.pytorch.org/whl/cu130}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]