portainer-core/stacks/models.yml

version: '3.8'

# Models Stack - GPU-Accelerated ML/AI Model Services
# Purpose: All GPU model serving (LLM inference, audio generation, 3D generation)
# Ports: 11434 (Ollama API), 11500 (Stable Audio UI), 11510 (TRELLIS UI)
# GPU: YES - Shared RTX 2080 Ti (11GB VRAM)
# Network: docker-dataplane

services:
  # ============================================
  # Ollama - LLM Inference Server
  # API: http://localhost:11434
  # ============================================
  ollama:
    image: ollama/ollama:latest
    container_name: ollama
    restart: unless-stopped
    ports:
      - "11434:11434"
    volumes:
      # Model storage on SSD for fast load times
      - /home/jpmschweitzer/docker-data/ollama/models:/root/.ollama
    environment:
      - TZ=Europe/Amsterdam
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=all
      # Process requests sequentially to avoid batch overflow panics
      - OLLAMA_NUM_PARALLEL=1
      # Unload LLMs after 5 minutes idle (keeps VRAM free for other services)
      - OLLAMA_KEEP_ALIVE=5m
      # Only keep one model loaded at a time (embedding model stays, LLMs swap)
      - OLLAMA_MAX_LOADED_MODELS=1
    healthcheck:
      test: ["CMD-SHELL", "curl -fSs http://localhost:11434/api/tags || exit 1"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s
    deploy:
      resources:
        limits:
          memory: 8G
        reservations:
          memory: 1G
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - docker-dataplane

  # ============================================
  # Stable Audio Open - AI Audio Generation
  # Web UI: http://localhost:11500
  # Image: locally built
  # ============================================
  stable-audio:
    image: stable-audio-open:local
    container_name: stable-audio
    restart: unless-stopped
    ports:
      - "11500:8000"
    volumes:
      # HuggingFace cache for model weights (~6GB)
      - /home/jpmschweitzer/docker-data/stable-audio/hf-cache:/root/.cache/huggingface
    environment:
      - TZ=Europe/Amsterdam
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
      - HF_TOKEN=${HF_TOKEN}
    healthcheck:
      test: ["CMD-SHELL", "curl -fSs http://localhost:8000/ || exit 1"]
      interval: 60s
      timeout: 30s
      retries: 3
      start_period: 300s
    deploy:
      resources:
        limits:
          memory: 16G
        reservations:
          memory: 8G
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    labels:
      - "com.centurylinklabs.watchtower.enable=false"
    networks:
      - docker-dataplane

  # ============================================
  # TRELLIS - 3D Model Generation (Low-VRAM Fork)
  # Web UI: http://localhost:11510
  # Image: locally built
  # ============================================
  trellis:
    image: trellis:local
    container_name: trellis
    restart: unless-stopped
    ports:
      - "11510:7860"
    volumes:
      # SSD: HuggingFace model cache (~5GB)
      - /home/jpmschweitzer/docker-data/trellis/hf-cache:/root/.cache/huggingface
      # HDD: Generated GLB files
      - /mnt/media/trellis/outputs:/app/outputs
      # Entrypoint script (patches Gradio bugs without rebuilding image)
      - /home/jpmschweitzer/docker-data/trellis/entrypoint.sh:/app/entrypoint.sh:ro
    command: ["bash", "/app/entrypoint.sh"]
    environment:
      - TZ=Europe/Amsterdam
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
      # VRAM Optimization (for 11GB card)
      - ATTN_BACKEND=xformers
      - SPCONV_ALGO=native
      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
      # Gradio must bind to 0.0.0.0 inside Docker
      - GRADIO_SERVER_NAME=0.0.0.0
      # HuggingFace token for model downloads (set in Portainer)
      - HF_TOKEN=${HF_TOKEN:-}
    healthcheck:
      test: ["CMD-SHELL", "curl -fSs http://localhost:7860/ || exit 1"]
      interval: 60s
      timeout: 30s
      retries: 3
      start_period: 300s
    memswap_limit: 20G
    deploy:
      resources:
        limits:
          memory: 20G
        reservations:
          memory: 4G
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    labels:
      - "com.centurylinklabs.watchtower.enable=false"
    networks:
      - docker-dataplane

networks:
  docker-dataplane:
    external: true
    name: docker-dataplane

# =============================================================================
# GPU CONTENTION
# =============================================================================
#
# All three services share the RTX 2080 Ti (11GB VRAM).
# Ollama unloads models after 5 min idle to free VRAM.
# For heavy Stable Audio or TRELLIS generations, consider stopping Ollama:
#   docker stop ollama && <generate> && docker start ollama
#
# =============================================================================
# OLLAMA
# =============================================================================
#
# After Deployment:
# 1. Verify GPU access: docker exec ollama nvidia-smi
# 2. Pull a model: docker exec ollama ollama pull llama3.2:3b
# 3. List models: docker exec ollama ollama list
# 4. Test inference: docker exec ollama ollama run llama3.2:3b "Hello"
#
# Recommended Models for RTX 2080 Ti (11GB VRAM):
# - llama3.2:3b (2GB) - Fast, general purpose
# - mistral:7b (4GB) - High quality, coding
# - codellama:7b (4GB) - Code-specialized
# - phi3:mini (2GB) - Fast reasoning
#
# =============================================================================
# STABLE AUDIO
# =============================================================================
#
# Prerequisites:
# 1. Build image: cd ~/docker-data/stable-audio-open && docker build -t stable-audio-open:local .
# 2. Create HuggingFace token: https://huggingface.co/settings/tokens (read access)
# 3. Accept model license: https://huggingface.co/stabilityai/stable-audio-open-1.0
# 4. mkdir -p ~/docker-data/stable-audio/hf-cache
# 5. Set HF_TOKEN in Portainer environment variables
#
# =============================================================================
# TRELLIS
# =============================================================================
#
# Prerequisites:
# 1. Build image: cd ~/docker-data/trellis && docker build -t trellis:local . (~20-30 min)
# 2. mkdir -p ~/docker-data/trellis/hf-cache /mnt/media/trellis/outputs
#
# Game Asset Pipeline:
# 1. Open Gradio UI at http://tower-of-joy:11510
# 2. Upload reference image or enter text prompt
# 3. Generate 3D model (uses 6-8GB VRAM)
# 4. Download GLB file from outputs
# 5. Import into Blender, apply textures, render sprites