# ============================================================================
# stt-server (whisper.cpp Vulkan + pyannote) — docker compose
# ----------------------------------------------------------------------------
# Uso:
#   export RENDER_GID=$(getent group render | cut -d: -f3)   # GID gruppo render
#   docker compose up -d --build
#
# GPU: AMD/Intel via Mesa RADV (/dev/dri/renderD128). Su host dove il node
# è world-writable (es. Fedora) group_add non è necessario ma innocuo.
#
# Variabili utili:
#   MODELS_DIR : dir con ggml-large-v3-turbo.bin + ggml-silero-v6.2.0.bin
#   HF_CACHE   : cache HuggingFace (pyannote Community-1, gated)
#   HF_TOKEN   : token HF (solo se la cache non contiene i modelli)
#   PORT       : porta host (default 8883)
# ============================================================================
services:
  stt-server:
    build:
      context: .
      dockerfile: Dockerfile
    image: stt-server:vulkan
    container_name: stt-server
    devices:
      - /dev/dri/renderD128:/dev/dri/renderD128
    group_add:
      - "${RENDER_GID:-44}"
    environment:
      GGML_BACKEND: Vulkan0
      WHISPER_CLI: /app/whisper-cli
      MODELS_DIR: /models
      STT_PYTHON: python3
      HF_HOME: /hf-cache
      HF_TOKEN: ${HF_TOKEN:-}
      PORT: "8883"
      GLOSSARY_FILE: /glossary/glossary.json
    volumes:
      - "${MODELS_DIR:-/home/enne2/dev/whisper.cpp/models}:/models:ro"
      - "${HF_CACHE:-/home/enne2/.cache/huggingface}:/hf-cache:ro"
      - "${GLOSSARY_FILE:-/home/enne2/dev/stt/glossary.json}:/glossary/glossary.json:ro"
    ports:
      - "${PORT:-8883}:8883"
    restart: unless-stopped