Skip to content

Docker

Containerized deploy on Linux. ASR model included in the build; local summary LLM downloaded automatically at startup if RAM ≥ 16 GB.

Profiles

Service Profile Base image
sbobinator-cpu cpu python:3.12-slim
sbobinator-gpu gpu nvidia/cuda:12.6

Build (~2.5 GB ASR model)

cd docker
docker compose --profile cpu build

During build:

  1. pip install -e ".[ui,summarize]" — FastAPI, NeMo, LLM providers, llama-cpp-python
  2. download_model.py → /models/parakeet-tdt-0.6b-v3.nemo

Startup

docker compose --profile cpu up

At startup (sbobina docker-ui):

  1. Checks container RAM (≥ 16 GB)
  2. If OK and Qwen is missing → downloads ~2 GB from HuggingFace
  3. Starts UI at http://localhost:8502 (default compose mapping 8502:8501)

First startup with Qwen download: may take 10–20 minutes (network). Logs show progress.

If RAM is insufficient: skips download; cloud summary remains available with an API key.

No need to run docker exec ... download_summary_llm.py manually anymore.

Volumes

volumes:
  - ../data:/data
  - sbobinator-qwen:/models/qwen2.5-3b-instruct
Host / volume Container Purpose
./data/input/ /data/input/ Source files
./data/output/ /data/output/ Jobs and results
./data/.secrets/ /data/.secrets/ Summary API keys
sbobinator-qwen (named) /models/qwen2.5-3b-instruct/ Qwen GGUF (persists across restarts)

Environment variables

environment:
  - NEMO_CACHE_DIR=/models
  - SBOBINATOR_DATA=/data
  - SBOBINATOR_UI_HOST=0.0.0.0
  # Cloud summary (optional):
  - SBOBINATOR_DEEPSEEK_API_KEY=sk-...

Or configure from the UI → saved in /data/.secrets/summary_keys.json.

GPU

docker compose --profile gpu up

Requires NVIDIA Container Toolkit.

Differences vs local Windows

Aspect Local Docker (cpu profile)
UI 127.0.0.1:8501 localhost:8502 → container 8501
ASR models/ in repo /models in image
Cloud summary API key in secrets Same (volume /data)
Local Qwen download_summary_llm.py Auto on Docker startup if RAM ≥ 16 GB
API SSL truststore (Windows) Usually not needed (Linux)

Production

  • Pre-build the image once
  • Mount only data/
  • Back up data/output/jobs/
  • Healthcheck on /partials/queue