Linux-Install-Scripts/docker-compose.yml at main · NextLVLHasH/Linux-Install-Scripts · 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
# Compose file for ZimaOS / CasaOS / generic Docker hosts.
# Import via the CasaOS dashboard ("Custom Install") or run:
#   docker compose up -d
# CasaOS one-liner install (per the user's note):
#   curl -fsSL https://get.casaos.io | sudo bash
#
# Persistent data lives in named volumes so model downloads + datasets survive
# container restarts. GPU support requires the NVIDIA Container Toolkit on the
# host. Comment out the `deploy` block if you only have a CPU.

services:
  trainer-dashboard:
    image: ml-stack-trainer-dashboard:latest
    build:
      context: .
      dockerfile: dashboard/Dockerfile
    container_name: trainer-dashboard
    restart: unless-stopped
    ports:
      - "8888:8888"
      - "1234:1234"
    environment:
      AUTO_LAUNCH_LMSTUDIO: "0"   # headless container; llama-server is the provider
      DASHBOARD_HOST: "0.0.0.0"
      DASHBOARD_PORT: "8888"
      LMS_API_PORT: "1234"
      LLAMA_BIND: "0.0.0.0"
      LLAMA_DIR: "/opt/llama.cpp-bin"
      LLAMA_BIN: "/opt/llama.cpp-bin/llama-server"
      GGUF_MODELS_DIR: "/app/models"
      HF_HOME: "/app/hf-cache"
    volumes:
      - models:/app/models
      - data:/app/data
      - runs:/app/runs
      - hf-cache:/app/hf-cache
    # Multi-GPU note: `count: all` exposes every visible NVIDIA GPU to the
    # container, which is what DDP/FSDP/ZeRO-3 need. To pin specific GPUs use
    # `device_ids: ['0','1','2','3']` instead of `count`. NVIDIA Container
    # Toolkit must be installed on the host. Comment the whole `deploy` block
    # for CPU-only operation.
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    # NCCL needs shared-mem >64MB and IPC=host for stable multi-GPU collectives.
    shm_size: "8gb"
    ipc: host
    ulimits:
      memlock: -1
      stack: 67108864
    labels:
      # CasaOS metadata (shows up in the CasaOS dashboard)
      casaos.icon: "https://cdn-icons-png.flaticon.com/512/8055/8055598.png"
      casaos.name: "ML Trainer"
      casaos.title: "ML Stack Trainer"
      casaos.category: "AI"
      casaos.port_map: "8765"

volumes:
  models:
  data:
  runs:
  hf-cache: