continuum/docker-compose.mac.yml at main · CambrianTech/continuum · 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
# Continuum — Mac override (continuum-core runs NATIVE, not containerized)
#
# Apple's hypervisor exposes no GPU to Linux containers — Docker themselves
# confirmed this in Feb 2026 ("Metal GPU access requires direct hardware
# access, no GPU passthrough for Metal in containers"). Trying to put
# continuum-core in a container on Mac means:
#   - Candle embeddings (AllMiniLM-L6-V2) fall back to CPU
#   - Bevy headless avatar render falls back to CPU
#   - Vision processing pipeline falls back to CPU
#   - Audio MPS paths (once wired) fall back to CPU
#   - LLM inference needs krunkit+MoltenVK which tops out at 80% of native
#     and fails to initialize on many M-series configs (ERROR_OUT_OF_HOST_
#     MEMORY at vkCreateInstance — verified 2026-04-16 on M5 Sequoia)
#
# The correct architecture for Mac:
#
#   continuum-core           → NATIVE process on host (Metal-linked cargo
#                              build). Launched via `npm start`, not docker.
#   LLM inference            → Docker Model Runner's vllm-metal backend,
#                              also running NATIVE on host (Docker Desktop
#                              orchestrates but compute happens on Apple
#                              GPU directly). continuum-core routes LLM
#                              calls to http://host.docker.internal:<port>/
#                              or localhost (since it's also on the host).
#   Support services         → IN CONTAINERS (this file):
#                              postgres, node-server, widget-server,
#                              livekit-bridge, model-init.
#
# This override sets continuum-core's `deploy.replicas: 0` so `docker compose
# up` on Mac brings up everything EXCEPT continuum-core. The install.sh
# Mac branch then runs `npm start` separately to launch continuum-core-server
# natively on the host.
#
# Usage (Mac — Carl and Dev both use this):
#   docker compose -f docker-compose.yml -f docker-compose.mac.yml pull
#   docker compose -f docker-compose.yml -f docker-compose.mac.yml up -d
#   cd src && npm start   # native continuum-core-server with Metal
#
# Linux (AMD/Intel/VirtIO) continues to use continuum-core-vulkan in a
# container — that path works because /dev/dri passthrough is real on
# Linux. See docker-compose.gpu.yml for the Nvidia+CUDA path.

services:

  # continuum-core does NOT run in a container on Mac. Runs natively via
  # `npm start` so it can link Metal for Candle embeddings + Bevy render +
  # vision + audio + direct llama.cpp ggml-metal inference.
  continuum-core:
    deploy:
      replicas: 0

  # Legacy inference service (upstream llama.cpp:server) stays off — same
  # reason as the gpu profile. Personas talk to native continuum-core's
  # scheduler + Docker Model Runner's vllm endpoint.
  inference:
    deploy:
      replicas: 0

  # node-server runs in Docker on Mac (containerized per Option B).
  # It needs to reach the NATIVE continuum-core-server on the host.
  # Unix sockets don't traverse Docker Desktop's VM boundary on Mac, so
  # continuum-core binds an additional TCP listener on 127.0.0.1:9100
  # (via CONTINUUM_CORE_TCP=9100 exported by install.sh before `npm start`),
  # and node-server connects to that via host.docker.internal (Docker
  # Desktop's standard alias for the host).
  #
  # No postgres override needed — postgres is now opt-in via the `postgres`
  # compose profile, and node-server no longer connects to any DB directly
  # (all data ops go through continuum-core via IPC with opaque handles).
  node-server:
    environment:
      # Added for Mac: overrides the default Unix-socket path lookup.
      - CONTINUUM_CORE_URL=tcp://host.docker.internal:9100