# GigaAM ASR — deployment via Docker Compose (self-hosted: Linux server / NAS / mini-PC).

#

# Run: `docker compose up -d --build` (a container-management UI works too). Before starting:

# 1) Create a `.env` file alongside (see .env.example) and set API_KEY.

# 2) Create the ./models directory and make UID 1000 its owner (chown 1000:1000 ./models),

# otherwise non-root in the container cannot write weights there. See README → "Deployment (Docker Compose)".

# The first start downloads the GigaAM weights into ./models (minutes) → then the healthcheck becomes healthy.

services:

gigaam-api:

build:

context: .

dockerfile: Dockerfile

image: gigaam-api:latest

container_name: gigaam-api

# required: false → without .env the service starts on defaults (config.py); for production

# create .env (see .env.example) and set API_KEY — it is picked up automatically.

env_file:

- path: .env

required: false

ports:

- "8000:8000"

volumes:

# The weights cache survives container re-creation; the first start downloads the weights here.

- ./models:/data/models

restart: unless-stopped

healthcheck:

# Without curl (absent in slim) — stdlib urllib, exec-form CMD. start_period is generous for

# the weights download on the first start: until they load uvicorn does not respond yet.

test: ["CMD", "python", "-c", "import sys, urllib.request; sys.exit(0 if urllib.request.urlopen('http://127.0.0.1:8000/health', timeout=5).status == 200 else 1)"]

interval: 30s

timeout: 10s

retries: 5

start_period: 600s

# --- Resource limits (a typical self-hosted host: ~4 cores / 8 GB) ---

# Tune to your hardware. The top-level mem_limit/cpus work more reliably

# (Compose specification v2) than deploy.resources (Swarm). Uncomment if needed:

# mem_limit: 6g

# cpus: 4.0

# Keep NUM_THREADS in .env ≤ the number of allocated cores (torch.set_num_threads).

# --- Optional weights warm-up (does NOT start with `up`; only via the `tools` profile) ---

# Pre-download the weights into ./models without bringing up the HTTP service (useful for a slow/offline

# host). Run: docker compose --profile tools run --rm download-weights

download-weights:

build:

context: .

dockerfile: Dockerfile

image: gigaam-api:latest

profiles: ["tools"]

env_file:

- path: .env

required: false

volumes:

- ./models:/data/models

command: ["python", "-m", "gigaam_api.download_weights"]

restart: "no"