docker-compose-strixhalo.yml

services:
  chatterbox-tts-server:
    build:
      context: .
      dockerfile: Dockerfile.strixhalo
    ports:
      - "${PORT:-8004}:8004"
    volumes:
      # Mount local config file for persistence
      - ./config.yaml:/app/config.yaml
      # Mount local directories for persistent app data
      - ./voices:/app/voices
      - ./reference_audio:/app/reference_audio
      - ./outputs:/app/outputs
      - ./logs:/app/logs
      - hf_cache:/app/hf_cache

    # --- ROCm GPU Access ---
    # Standard ROCm device access - required for AMD GPU acceleration
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - video
      - render
    ipc: host
    shm_size: 8g
    security_opt:
      - seccomp=unconfined
    
    # --- Optional: Enhanced ROCm Access ---
    # Uncomment the lines below if you experience GPU access issues
    # privileged: true
    # cap_add:
    #   - SYS_PTRACE
    # devices:
    #   - /dev/mem
    
    restart: unless-stopped
    environment:
      # Enable faster Hugging Face downloads
      - HF_HUB_ENABLE_HF_TRANSFER=1
      - HF_TOKEN=YOUR_TOKEN_HERE
      - TTS_ENGINE_DEVICE=cuda
      # Enable bfloat16 for T3/S3Gen — halves memory bandwidth on token generation
      # Set TTS_BF16=off if you experience precision issues
      - TTS_BF16=on
      - TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
      - MIOPEN_FIND_MODE=FAST
      - HSA_OVERRIDE_GFX_VERSION=11.0.0
      # Critical for Strix Halo unified memory (gfx1151): enables page-fault retry
      - HSA_XNACK=1
      # Reduces memory fragmentation on large unified memory pool
      - PYTORCH_ALLOC_CONF=expandable_segments:True
      # NOTE: HSA_OVERRIDE_GFX_VERSION should only be set by users with unsupported GPUs
      # Example usage: HSA_OVERRIDE_GFX_VERSION=10.3.0 docker compose up
      # Common values: 10.3.0 (RX 5000/6000), 11.0.0 (RX 7000), 9.0.6 (Vega)

volumes:
  hf_cache: