docker-compose.yml

# localplaud — one compose file, three hardware profiles. Pick ONE per host:
#
#   docker compose --profile cpu up -d --build   # CPU box / cloud VM — use a cloud ASR provider
#   docker compose --profile gpu up -d --build   # NVIDIA host — local faster-whisper on CUDA
#   docker compose --profile gpu --profile ollama up -d --build
#                                                # NVIDIA ASR + optional local text generation
#   docker compose --profile mac up -d --build   # Apple Silicon — see the note below
#
# Before first start:
#   cp .env.example .env                 # DOMAIN, ACME_EMAIL, Plaud cookie, API keys
#   cp config.example.toml config.toml   # pipeline / provider settings
#
# Caddy terminates HTTPS for $DOMAIN (automatic Let's Encrypt certificates)
# and proxies to whichever app service is active — they all share the network
# alias "app" on port 8080, and only one app profile runs at a time.
#
# macOS note: containers cannot reach the Metal GPU, so the `mac` profile is
# meant to run with a cloud ASR provider (set LOCALPLAUD_ASR__PROVIDER in .env).
# For on-device Metal ASR (mlx-whisper / whisper.cpp), run the app natively:
#     uv run localplaud run
# and use compose only for Caddy (`docker compose --profile mac up -d caddy`,
# pointing the Caddyfile at host.docker.internal:8080 instead of app:8080).
# In-container, host services like Ollama are reachable at
# http://host.docker.internal:11434.

x-localplaud: &localplaud
  env_file: .env
  environment:
    HF_TOKEN: ${LOCALPLAUD_DIARIZE__HF_TOKEN:-}
  volumes:
    - ./data:/app/data
    - ./config.toml:/app/config.toml:ro
    - model_cache:/home/localplaud/.cache
  expose:
    - "8080"
  ports:
    - "127.0.0.1:8081:8080"
  restart: unless-stopped
  networks:
    default:
      aliases:
        - app

services:
  localplaud-cpu:
    <<: *localplaud
    profiles: ["cpu"]
    build:
      context: .
      dockerfile: Dockerfile
      # Slim, cloud-only image: no faster-whisper/torch (pointless on a CPU box
      # and very slow to build on weak ARM). Use cloud ASR/LLM/embeddings —
      # set the providers + API keys in .env / config.toml.
      args:
        EXTRAS: "cloud"

  localplaud-gpu:
    <<: *localplaud
    profiles: ["gpu"]
    build:
      context: .
      dockerfile: Dockerfile.cuda
    gpus: all
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  localplaud-mac:
    <<: *localplaud
    profiles: ["mac"]
    build:
      context: .
      dockerfile: Dockerfile
    extra_hosts:
      - "host.docker.internal:host-gateway"

  ollama-gpu:
    image: ollama/ollama:0.31.2
    profiles: ["ollama"]
    environment:
      OLLAMA_KEEP_ALIVE: "2m"
      OLLAMA_CONTEXT_LENGTH: "8192"
    volumes:
      - ollama_data:/root/.ollama
    expose:
      - "11434"
    gpus: all
    restart: unless-stopped

  caddy:
    image: caddy:2
    profiles: ["cpu", "gpu", "mac"]
    ports:
      - "80:80"
      - "443:443"
    env_file: .env
    volumes:
      - ./Caddyfile:/etc/caddy/Caddyfile:ro
      - caddy_data:/data
      - caddy_config:/config
    # Loose dependencies: only the app service of the active profile starts;
    # the others are skipped (required: false).
    depends_on:
      localplaud-cpu:
        condition: service_started
        required: false
      localplaud-gpu:
        condition: service_started
        required: false
      localplaud-mac:
        condition: service_started
        required: false
    restart: unless-stopped

volumes:
  caddy_data:
  caddy_config:
  model_cache:
  ollama_data: