compose.yml

name: fish-speech

services:
  # OpenAI-compatible TTS API (streaming, emotion tags, voice cloning,
  # multi-speaker via voice_map). This is the DEFAULT service:
  #   cp .env.example .env  &&  docker compose up -d --build
  # Configure everything in .env (port, GPU id, quant level, voices folder, …).
  # Requires the NVIDIA Container Toolkit on the host.
  fish-openai:
    build:
      context: .
      dockerfile: docker/Dockerfile
      target: openai
      args:
        BACKEND: cuda
        UV_EXTRA: "${FISH_UV_EXTRA:-cu128}"      # cu128 = torch 2.8 + CUDA 12.8
        CUDA_VER: "${FISH_CUDA_VER:-12.8.1}"
        UV_VERSION: "${UV_VERSION:-0.8.15}"
        USER_UID: "${FISH_UID:-1000}"
        USER_GID: "${FISH_GID:-1000}"
    image: fish-speech-openai:cu128
    container_name: fish-openai
    restart: unless-stopped
    ports:
      - "${FISH_HOST_PORT:-8770}:8770"
    environment:
      FISH_DEVICE: "cuda:0"                       # in-container the chosen GPU is cuda:0
      FISH_LISTEN: "0.0.0.0:8770"
      FISH_QUANTIZE: "${FISH_QUANTIZE:-none}"     # none | int8 | int4
      # Load pre-quantized weights (tools/quantize_save.py) to skip per-boot
      # re-quant. Path is inside the container (checkpoints is mounted). Empty
      # = quantize live each boot.
      FISH_QUANTIZED_WEIGHTS: "${FISH_QUANTIZED_WEIGHTS:-}"
      FISH_QUANT_GROUPSIZE: "${FISH_QUANT_GROUPSIZE:-128}"
      FISH_COMPILE: "${FISH_COMPILE:-1}"
      FISH_HALF: "${FISH_HALF:-1}"
      FISH_MAX_SEQ_LEN: "${FISH_MAX_SEQ_LEN:-0}"
      FISH_API_KEY: "${FISH_API_KEY:-}"
      VOICES_DIR: "/app/voices"                   # auto-register + watch (mounted below)
      FISH_VOICES_SCAN_INTERVAL: "${FISH_VOICES_SCAN_INTERVAL:-30}"
      FISH_AUTO_TRANSCRIBE: "${FISH_AUTO_TRANSCRIBE:-1}"  # ASR a transcript for voices that lack one
      FISH_ASR_MODEL: "${FISH_ASR_MODEL:-small}"  # faster-whisper model (small=multilingual)
      FISH_ASR_DEVICE: "${FISH_ASR_DEVICE:-cpu}"  # cpu | cuda:0 (cpu avoids VRAM contention)
      FISH_CONCURRENCY: "${FISH_CONCURRENCY:-1}"
      FISH_QUEUE_TIMEOUT: "${FISH_QUEUE_TIMEOUT:-300}"
      # Persist torch.compile/Triton kernels to the mounted cache (below) so a
      # restart skips the ~4-min CPU kernel recompile.
      TORCHINDUCTOR_CACHE_DIR: "/app/.torch-cache/inductor"
      TRITON_CACHE_DIR: "/app/.torch-cache/triton"
      TORCHINDUCTOR_FX_GRAPH_CACHE: "1"
      # Incremental decode for streaming: first audio in ~1.5s vs ~12s on long
      # turns (0 = off). Only affects streaming requests; generation unchanged.
      FISH_STREAM_CHUNK_TOKENS: "${FISH_STREAM_CHUNK_TOKENS:-32}"
    volumes:
      - "${FISH_CHECKPOINTS_HOST_DIR:-./checkpoints}:/app/checkpoints"
      - "${FISH_VOICES_HOST_DIR:-./voices}:/app/voices"
      - "${FISH_REFERENCES_HOST_DIR:-./references}:/app/references"
      # Persist torch.compile (Inductor) + Triton kernel caches so a restart
      # skips the ~4-min CPU recompile. Pre-create the host dir owned by
      # FISH_UID:FISH_GID so the container (non-root) can write it.
      - "${FISH_TORCH_CACHE_HOST_DIR:-./.torch-cache}:/app/.torch-cache"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["${FISH_GPU_ID:-0}"]   # physical GPU id on the host
              capabilities: [gpu]

  # --- Upstream fishaudio services (opt-in via profiles) ---
  webui:
    extends:
      file: compose.base.yml
      service: app-base
    build:
      target: webui
    environment:
      COMPILE: ${COMPILE:-0}
    profiles: ["webui"]
    ports:
      - "${GRADIO_PORT:-7860}:7860"

  server:
    extends:
      file: compose.base.yml
      service: app-base
    build:
      target: server
    environment:
      COMPILE: ${COMPILE:-0}
    profiles: ["server"]
    ports:
      - "${API_PORT:-8080}:8080"