Skip to content

INTERNAL_OPERATIONS.md

Latest commit

 

History

History
65 lines (49 loc) · 2.69 KB

INTERNAL_OPERATIONS.md

File metadata and controls

65 lines (49 loc) · 2.69 KB
Raw

Internal operations (docker-agent and related)

Internal, host-specific notes. Not intended as public deployment documentation.

docker-agent layout

  • Host: shared live host; multiple containers coexist.

  • Docker Compose lives at ~/docker/docker-agent/docker-compose.yml (root@docker-agent, 192.168.10.72).

  • Qwen3-TTS data lives under:

    • /var/data/autopirate/persona-forge/model → HF cache (MODEL_CACHE_PATH)
    • /var/data/autopirate/persona-forge/openvino → OpenVINO IR (OV_DATA_PATH)
    • /var/data/autopirate/persona-forge/voices → VoiceDesign voices (VOICE_LIBRARY_PATH)
    • /var/data/autopirate/persona-forge/voice/ → Reference audio (e.g. voice_A.wav)

Basic operations

  • Start:
    • docker compose -f ~/docker/docker-agent/docker-compose.yml up -d persona-forge
  • Stop (don’t touch unrelated services):
    • docker compose -f ~/docker/docker-agent/docker-compose.yml down persona-forge
  • Pull new image (example):
    • export PERSONA_FORGE_IMAGE=ghcr.io/nmorgowicz-org/persona-forge:<sha>
    • docker compose -f ~/docker/docker-agent/docker-compose.yml pull persona-forge
    • docker compose -f ~/docker/docker-agent/docker-compose.yml up -d persona-forge
  • Export (stop service first if constrained):
    • docker compose -f ~/docker/docker-agent/docker-compose.yml down persona-forge
    • docker compose -f ~/docker/docker-agent/docker-compose.yml run --rm export
    • docker compose -f ~/docker/docker-agent/docker-compose.yml up -d persona-forge

Never run a 13 GiB export alongside a 10 GiB serve on this box (OOM).

Validation after changes

After image, model, IR, or runtime-setting changes:

  • curl -fsS http://localhost:8318/health | python -m json.tool
  • docker inspect --format '{{.Image}}' persona-forge
  • docker exec persona-forge cat /sys/fs/cgroup/memory.current
  • docker exec persona-forge cat /sys/fs/cgroup/memory.peak
  • docker stats --no-stream persona-forge

Record:

  • Image tag/digest
  • Source commit
  • Model revision, IR metadata hash (from /health)
  • Backend, compression, cache profile
  • Prompts and sampling settings
  • Latency, audio duration/RTF
  • Memory current/peak, host available RAM, swap delta
  • Listening notes

Use deterministic greedy runs for code/parity comparisons; use production sampling for final listening and performance decisions.

Rollback

  • Fastest: restore previous immutable image and restart Compose.
  • Backend: TTS_BACKEND=pytorch; verify via /health. Slower, for recovery only.

Notes

  • Only touch persona-forge; other services (litellm*, headroom-proxy, crowdsec, hermes-*, *arr, searxng) must not be affected.
  • Keep port 8318 trusted-network-only or behind an authenticated reverse proxy.