Video-to-Gaussian/docker-compose.consolidated.yml at main · DreamLab-AI/Video-to-Gaussian · 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
102
services:
  gaussian-toolkit:
    build:
      context: .
      dockerfile: Dockerfile.consolidated
    image: gaussian-toolkit:latest
    container_name: gaussian-toolkit
    hostname: gaussian-toolkit
    runtime: nvidia

    env_file:
      - .env

    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility,graphics
      - CUDA_HOME=/usr/local/cuda
      - DISPLAY=:1
      - HF_TOKEN=${HF_TOKEN}
      - HF_HOME=/opt/hf-cache
      - ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY:-}
      - SAM3_BPE_PATH=/opt/sam3-repo/sam3/assets/bpe_simple_vocab_16e6.txt.gz
      - MILO_CONTAINER=milo

    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['0']
              capabilities: [gpu, compute, utility, graphics]
        limits:
          memory: 200g

    shm_size: "64g"

    ports:
      - "7860:7860"    # Web interface (Flask)
      - "7681:7681"    # Web terminal (ttyd)
      - "8188:8188"    # ComfyUI
      - "45677:45677"  # LichtFeld Studio
      - "5901:5901"    # VNC

    volumes:
      # Model staging — pre-populate this with downloaded models
      - /home/john/comfyui-models-staging:/models-staging:ro
      # Persistent output data (web UI writes to /data/output)
      - ./output:/data/output
      # Host-compiled LichtFeld binary (bind-mount from host build)
      - /home/john/githubs/gaussian/LichtFeld-Studio/build:/opt/gaussian-toolkit/build:ro
      # Host CUDA libs (if container CUDA version differs from host)
      - /opt/cuda/lib64:/opt/host-cuda-libs:ro
      # Persistent model cache (survives rebuilds)
      - models-data:/opt/models
      # HuggingFace cache (shared path via HF_HOME env var)
      - hf-cache:/opt/hf-cache
      # Claude Code OAuth session (survives container recreation)
      - claude-session:/home/ubuntu/.claude
      # Live pipeline source (override baked-in code for dev iteration)
      - ./src/pipeline:/opt/gaussian-toolkit/src/pipeline
      - ./src/web:/opt/gaussian-toolkit/src/web
      - ./CLAUDE_CONTAINER.md:/opt/gaussian-toolkit/CLAUDE.md:ro

    restart: unless-stopped

    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:7860/health"]
      interval: 30s
      timeout: 10s
      retries: 5
      start_period: 120s

  # MILo sidecar — high-quality mesh extraction (Ubuntu 22.04 + CUDA 11.8)
  # Build: docker compose build milo
  # The main container calls MILo via: docker exec milo python3 train.py ...
  milo:
    build:
      context: .
      dockerfile: docker/Dockerfile.milo
    image: gaussian-toolkit-milo:latest
    container_name: milo
    runtime: nvidia
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['1']
              capabilities: [gpu, compute, utility]
    volumes:
      - ./output:/data/output
      - hf-cache:/opt/hf-cache
    entrypoint: ["sleep", "infinity"]
    restart: unless-stopped

volumes:
  models-data:
    driver: local
  hf-cache:
    driver: local
  claude-session:
    driver: local