protoVoice/docker-compose.yml at main · protoLabsAI/protoVoice · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
# protoVoice — one-command deploy
#
#   docker compose up -d
#
# Services:
#   - protovoice          (GPU 0)   WebRTC UI + Whisper + Kokoro + routing vLLM
#   - fish-speech         (GPU 1)   Fish Audio S2-Pro TTS sidecar — default backend
#   - fish-openai-shim              OpenAI /v1/audio/speech ↔ Fish /v1/tts proxy on :8093
#                                   (no GPU — pure HTTP + ffmpeg; talks to fish-speech)
#
# Use `TTS_BACKEND=kokoro docker compose up protovoice` to run without Fish.

services:
  fish-speech:
    # Built from ~/dev/fish-speech checkout — has the .venv + checkpoints.
    build:
      context: ../fish-speech
      dockerfile: ${PWD}/Dockerfile.fish
    container_name: protovoice-fish
    restart: unless-stopped
    runtime: nvidia
    ports:
      - "${FISH_PORT:-8092}:8092"
    volumes:
      - ${FISH_REFERENCES_DIR:-/mnt/data/fish-references}:/app/references
      # The Fish Dockerfile COPYs everything, but .dockerignore excludes
      # .venv/ and the checkpoints (too large) don't belong baked into the
      # image. Bind-mount both from the sibling ../fish-speech checkout that
      # the `build.context` above already targets.
      - ${FISH_SPEECH_DIR:-../fish-speech}/.venv:/app/.venv:ro
      - ${FISH_SPEECH_DIR:-../fish-speech}/checkpoints:/app/checkpoints:ro
    environment:
      - NVIDIA_VISIBLE_DEVICES=${FISH_GPU:-1}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["${FISH_GPU:-1}"]
              capabilities: [gpu]
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8092/v1/health')"]
      interval: 30s
      timeout: 10s
      start_period: 600s   # torch.compile warmup can take ~3-5 min from cold
      retries: 5

  fish-openai-shim:
    build:
      context: ./services/fish-openai-shim
    container_name: protovoice-fish-shim
    restart: unless-stopped
    depends_on:
      fish-speech:
        condition: service_started
    ports:
      - "${FISH_SHIM_PORT:-8093}:8093"
    environment:
      - FISH_URL=http://fish-speech:8092
      - FISH_SAMPLE_RATE=44100

  protovoice:
    build: .
    container_name: protovoice
    restart: unless-stopped
    runtime: nvidia
    depends_on:
      # Only required when TTS_BACKEND=fish (the default).
      fish-speech:
        condition: service_started
    ports:
      - "${PORT:-7866}:7866"
    volumes:
      - ${HF_HOME:-/mnt/models/huggingface}:/models
    # env_file picks up secrets like WORKSTACEAN_API_KEY / AVA_API_KEY that
    # aren't enumerated below — don't require a compose edit for every new
    # delegate key.
    env_file:
      - path: .env
        required: false
    environment:
      - NVIDIA_VISIBLE_DEVICES=${PROTOVOICE_GPU:-0}
      - HF_HOME=/models
      - PORT=7866
      - VLLM_PORT=8100
      - LLM_MODEL=${LLM_MODEL:-Qwen/Qwen3.5-4B}
      - WHISPER_MODEL=${WHISPER_MODEL:-openai/whisper-large-v3-turbo}
      - TTS_BACKEND=${TTS_BACKEND:-fish}
      - FISH_URL=${FISH_URL:-http://fish-speech:8092}
      - FISH_REFERENCE_ID=${FISH_REFERENCE_ID:-}
      - KOKORO_VOICE=${KOKORO_VOICE:-af_heart}
      - SYSTEM_PROMPT=${SYSTEM_PROMPT:-}
      - START_VLLM=${START_VLLM:-1}
      - LLM_URL=${LLM_URL:-}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["${PROTOVOICE_GPU:-0}"]
              capabilities: [gpu]