docker

HARRIER Docker Matrix

HARRIER ships three container variants. They all install the package from the same pyproject.toml, just with different base images and entrypoints.

Image	Base	Purpose	Backend
robotflowlabs/harrier-cuda	anima-base:jazzy	Training + serving on the GPU server	cuda
robotflowlabs/harrier-mlx	anima-base-mlx:jazzy	Apple Silicon development + inference	mlx
robotflowlabs/harrier-serve	anima-serve:jazzy	Production serving (no training deps)	cuda / cpu

Build

# From the repo root
docker compose -f docker/docker-compose.yml build harrier-cuda
docker compose -f docker/docker-compose.yml --profile mlx build harrier-mlx
docker compose -f docker-compose.serve.yml build harrier-serve

Run

docker compose -f docker-compose.serve.yml up -d
curl http://localhost:8010/health
curl http://localhost:8010/ready

Weights are mounted read-only from /mnt/artifacts-datai/models/project_harrier. That is the single source of truth for trained artifacts (see .claude/rules/save_checkpoint.md).

Environment variables

All three images accept:

Variable	Purpose
HARRIER_WEIGHTS	Absolute path to best.pt inside the container.
HARRIER_BACKEND	auto / cuda / mlx / cpu.
ANIMA_BACKEND	Legacy alias honoured by anima_harrier.device.

Healthcheck

Dockerfile.serve ships with a built-in HTTP healthcheck hitting /health. Compose uses the same probe so the orchestrator can tell the difference between "container alive" and "service reachable".