The Reef image bundles the full runtime (Slime, SGLang, Ray, Megatron) so that a deployment runs entirely inside one image as plain processes. There is no docker-compose layer: reef serve -c <config> reads the YAML's services list and starts each declared process directly, with PID/log files under /tmp/reef-stack.

docker build -f docker/Dockerfile.reef -t reef .

The image inherits the GPU training stack and installs Reef's exact runtime pin from pyproject.toml, together with SGLang, Ray, and the training dependencies. Having the runtime in the image does not mean its training driver is running — only the services listed in your config run.

The default image pins the SGLang revision carrying Reef's adapter receiver, so LoRA training (--megatron-lora-rank) works on any weight-training recipe without a separate image. Override the revision with --build-arg SGLANG_COMMIT=<sha>.

TTT-Discover's Qwen3-8B LoRA experiment uses the optional tttd target. It adds only the Erdős evaluator's solver dependencies for generated programs, alongside a qualified Slime digest:

docker build -f docker/Dockerfile.reef --target tttd \
  --build-arg SLIME_IMAGE_TAG='latest@sha256:a97ec147e37bef050337a9b229036eda00b4aa9c4d02b31a0109dc850f8ca342' \
  -t reef-tttd:qwen3-8b .

Each config declares its services declaratively (name, command, ready probe, depends_on). Commands use ${a.b.c} interpolation against the rest of the config, so adding or changing a service is a YAML-only edit. The orchestrator launches Reef's HTTP child internally with the same config; all other services are plain shell commands. See Evolve your model for the training flow and data contract.

Inside the image (or any host with the deps installed):

export REEF_TOKEN=$(openssl rand -hex 16)
# edit the stack yaml: set model paths / provider creds
reef serve -c recipes/basic/local-sglang.yaml

Logs and PIDs land in /tmp/reef-stack/. To run just the reef HTTP service against an already-running provider, use a config whose services list contains only Reef:

reef serve -c recipes/basic/external-provider.yaml

Agent records and exported checkpoints persist under reef.state_dir (/var/lib/reef by default). Stop the processes (kill the PIDs in /tmp/reef-stack/) but keep state; to also delete recorded agent records, remove that directory.