Intelligent-Video-Management-System/docker-compose.gpu.yml at main · VariPhiGen/Intelligent-Video-Management-System · 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
# GPU overlay (legacy nvidia runtime) — layered by ./vms only when the daemon
# reports an `nvidia` runtime but no CDI spec; hosts with CDI get
# docker-compose.gpu-cdi.yml instead, which survives `systemctl daemon-reload`
# (this hook-based path silently loses GPU access on reload — NVML "Unknown
# Error" — until the container is recreated).
# Never merge this into docker-compose.yml: a device reservation in
# the base file makes `api` refuse to START on an appliance without the NVIDIA
# container toolkit, and this product ships to CPU-only boxes too. Conditional
# composition is the same trick docker-compose.bridge.yml uses for networking.
#
# Grants the API metrics-only GPU visibility so nvidia-smi is mounted in and
# the topbar gauge has something to read. Inference lives in its own
# containers; this one must never hold a compute context.
#
#   ./vms up -d              # auto-detected
#   VMS_GPU=off ./vms up -d  # force off (troubleshooting)

services:
  api:
    environment:
      NVIDIA_VISIBLE_DEVICES: all
      # `utility` = nvidia-smi + NVML only. Deliberately NOT `compute`: the API
      # reads utilization, it does not run CUDA, and asking for compute would
      # pin driver context this process has no use for.
      NVIDIA_DRIVER_CAPABILITIES: utility
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu, utility]

  # NVR gets the VIDEO capability: HEVC→H.264 clip transcodes run on the GPU's
  # NVENC/NVDEC blocks (h264_nvenc) instead of pinning CPU cores with libx264.
  # The NVR probes with a real test encode at startup and falls back to the
  # CPU path when the probe fails, so this grant is safe on any driver state.
  nvr:
    environment:
      NVIDIA_VISIBLE_DEVICES: all
      NVIDIA_DRIVER_CAPABILITIES: compute,video,utility
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu, video, utility]

  # Smart Search: the one container that actually runs inference. Unlike api and
  # nvr above this opens a CUDA context and keeps it — that is the point.
  # Requires the image built with TORCH_VARIANT=cu121; a CPU build ignores the
  # device and logs that it is running on CPU.
  # Analytics: THE detector since 9f5a5ec — see docker-compose.gpu-cdi.yml for
  # why this block has to exist separately from smartsearch's.
  analytics:
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              count: all

  smartsearch:
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              capabilities: [gpu]
              count: all