compose.sglang.yaml

services:
  model:
    image: ${SGLANG_IMAGE:-lmsysorg/sglang:v0.5.17}
    entrypoint:
      - python3
      - /opt/frontier-agent/sglang_entrypoint.py
    environment:
      HF_TOKEN: ${HF_TOKEN:-}
      HUGGING_FACE_HUB_TOKEN: ${HF_TOKEN:-}
      SGLANG_MODEL_ID: ${SGLANG_MODEL_ID:-}
      # A non-empty host path is mounted read-only below and translated to
      # this stable container path. It takes precedence over SGLANG_MODEL_ID.
      SGLANG_LOCAL_MODEL_PATH: ${SGLANG_LOCAL_MODEL_PATH:+/opt/frontier-agent/local-model}
      SGLANG_SERVED_MODEL_NAME: ${SGLANG_SERVED_MODEL_NAME:-local-model}
      SGLANG_TP_SIZE: ${SGLANG_TP_SIZE:-1}
      SGLANG_CONTEXT_LENGTH: ${SGLANG_CONTEXT_LENGTH:-32768}
      SGLANG_MEM_FRACTION_STATIC: ${SGLANG_MEM_FRACTION_STATIC:-0.88}
      SGLANG_TOOL_CALL_PARSER: ${SGLANG_TOOL_CALL_PARSER:-}
      SGLANG_REASONING_PARSER: ${SGLANG_REASONING_PARSER:-}
      SGLANG_CHAT_TEMPLATE: ${SGLANG_CHAT_TEMPLATE:-}
      SGLANG_DTYPE: ${SGLANG_DTYPE:-auto}
      SGLANG_QUANTIZATION: ${SGLANG_QUANTIZATION:-}
      SGLANG_TRUST_REMOTE_CODE: ${SGLANG_TRUST_REMOTE_CODE:-0}
      SGLANG_EXTRA_ARGS: ${SGLANG_EXTRA_ARGS:-}
    volumes:
      - ./docker/sglang_entrypoint.py:/opt/frontier-agent/sglang_entrypoint.py:ro
      - ./docker/smoke_sglang.py:/opt/frontier-agent/smoke_sglang.py:ro
      - type: bind
        source: ${SGLANG_LOCAL_MODEL_PATH:-./docker/empty-model}
        target: /opt/frontier-agent/local-model
        read_only: true
      - huggingface-cache:/root/.cache/huggingface
    # Loopback only. The endpoint has no authentication and the agent reaches
    # it over the compose network, so publishing it on every interface would
    # hand anyone who can route to this host free inference on gated weights —
    # and these are exactly the cloud GPU boxes with a public address.
    ports:
      - 127.0.0.1:${SGLANG_PORT:-30000}:30000
    shm_size: 32gb
    ipc: host
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: ${SGLANG_GPU_COUNT:-1}
              capabilities: [gpu]
    healthcheck:
      test:
        - CMD
        - python3
        - -c
        - import urllib.request; urllib.request.urlopen('http://127.0.0.1:30000/health', timeout=5)
      interval: 10s
      timeout: 6s
      retries: 360
      start_period: 30s
    # Local TUI inference is opt-in and should not unexpectedly consume the GPU
    # after a host reboot. Operators can explicitly select unless-stopped.
    restart: ${SGLANG_RESTART_POLICY:-no}
    logging:
      driver: local
      options:
        max-size: 20m
        max-file: "3"

  agent:
    depends_on:
      model:
        condition: service_healthy
    # Merges with the agent service in compose.yaml, so APODEX_IN_CONTAINER
    # and SANDBOX_BACKEND=container carry over from there.
    environment:
      OPENAI_PROVIDER: local
      OPENAI_API_KEY: EMPTY
      OPENAI_BASE_URL: http://model:30000/v1
      OPENAI_MODEL: ${SGLANG_SERVED_MODEL_NAME:-local-model}
      OPENAI_CONTEXT_WINDOW: ${SGLANG_CONTEXT_LENGTH:-32768}
      # Two invariants, both checked by the doctors and by
      # test_deployment_config.py, so these fallbacks must satisfy them for an
      # env file that sets only a model: input + output <= context, and input
      # above 0.8 x OPENAI_CONTEXT_WINDOW. The latter is where tiered
      # compaction triggers, and ContextSizeGuard is critical, so a lower
      # budget aborts the run with budget_exhausted before compaction can ever
      # relieve it. These match .env.sglang.example and both launcher doctors.
      OPENAI_MAX_INPUT_TOKENS: ${SGLANG_MAX_INPUT_TOKENS:-27000}
      OPENAI_MAX_TOKENS: ${SGLANG_MAX_OUTPUT_TOKENS:-4096}
      SERPER_API_KEY: ${SERPER_API_KEY:-}
      SERPER_BASE_URL: ${SERPER_BASE_URL:-https://google.serper.dev}
      JINA_API_KEY: ${JINA_API_KEY:-}
      JINA_BASE_URL: ${JINA_BASE_URL:-https://r.jina.ai}

volumes:
  huggingface-cache: