docker-compose.nvidia.yml

Rama main · réplica de lectura del repositorio original.

services:
  # `count: all` goes through the host's CDI (Container Device
  # Interface) spec. it's written once and goes stale when you add
  # a card, then only hands over some of them. start.sh counts the
  # GPUs itself and asks for that many, taking the legacy path.
  ollama:
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: ${NVIDIA_GPU_COUNT:-all}
              capabilities: [gpu]
    environment:
      - CUDA_VISIBLE_DEVICES
      - OLLAMA_SCHED_SPREAD

  llamacpp:
    image: ghcr.io/ggml-org/llama.cpp:server-cuda@sha256:490d8cd8000203a3ab8fdfce14ef931e2aa85affd4dd4447bcc75f299f00fd9a
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: ${NVIDIA_GPU_COUNT:-all}
              capabilities: [gpu]
    environment:
      - CUDA_VISIBLE_DEVICES
      - LLAMA_ARG_SPLIT_MODE
      - LLAMA_ARG_TENSOR_SPLIT
      - LLAMA_ARG_MAIN_GPU
  # only the karaoke sidecar gets a GPU. the voice sidecar stays
  # CPU-only by design, see docker/tts.Dockerfile
  karaoke:
    build:
      args:
        # cu124 (torch built for CUDA 12.4) is gone. every +cu124
        # torch pins nvidia-cudnn-cu12==9.1.0.70 and that wheel is
        # off the index now, so the build dies with a wall of
        # resolver output that reads like torch itself is missing.
        # Override with KARAOKE_TORCH_INDEX if your driver is too old
        # for cu128.
        TORCH_INDEX: "${KARAOKE_TORCH_INDEX:-https://download.pytorch.org/whl/cu128}"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]