# Qwen3.8-27B — production stack: two GPU replicas + load balancer.
#
#   scripts/up.sh              bring everything up
#   scripts/status.sh          health, GPU, current SPEC/CTX per replica
#   scripts/ab-phase.sh phase1 crossover A/B (see README)
#
# Every knob lives in .env next to this file. Each replica is configured
# INDEPENDENTLY (GPU0_SPEC/GPU0_CTX vs GPU1_SPEC/GPU1_CTX) so the A/B crossover
# is an .env edit + `up.sh`, never a hand-edited docker run.
name: qwen38

x-replica: &replica
  image: qwen38-27b-3090:latest
  build:
    context: ${REPO_DIR:-./qwen38-rtx3090}
  ipc: host
  command: single
  restart: unless-stopped
  stop_grace_period: 30s
  volumes:
    - ${REPO_DIR:-./qwen38-rtx3090}/models:/app/models
    - qwen-cache:/cache
  networks: [qwen]
  # start_period is long on purpose: a COLD cache pays torch.compile + CUDA graph
  # capture + FlashInfer JIT (~3 min, up to ~15 on a first-ever run). Warm it is ~60 s.
  healthcheck:
    interval: 30s
    timeout: 5s
    retries: 3
    start_period: 900s

services:
  # Idempotent model prep — a no-op once models/ is populated (a few seconds).
  # Kept as a dependency so a half-prepared model dir can never be served.
  prepare:
    image: qwen38-27b-3090:latest
    build:
      context: ${REPO_DIR:-./qwen38-rtx3090}
    command: prepare
    ipc: host
    environment:
      HOME: /cache
    volumes:
      - ${REPO_DIR:-./qwen38-rtx3090}/models:/app/models
      - qwen-cache:/cache
    networks: [qwen]

  replica-gpu0:
    <<: *replica
    container_name: qwen38-gpu0
    environment:
      HOME: /cache
      PORT: "18020"
      VLLM_API_KEY: ${VLLM_API_KEY:?set VLLM_API_KEY in .env}
      SPEC: ${GPU0_SPEC:-}
      CTX: ${GPU0_CTX:-fast}
      PREFIX_CACHE: ${GPU0_PREFIX_CACHE:-0}
    ports:
      - "18020:18020"          # direct access + telemetry scrape target
    healthcheck:
      test: ["CMD", "curl", "-sf", "http://127.0.0.1:18020/health"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 900s
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["0"]
              capabilities: [gpu]
    depends_on:
      prepare:
        condition: service_completed_successfully

  replica-gpu1:
    <<: *replica
    container_name: qwen38-gpu1
    environment:
      HOME: /cache
      PORT: "18021"
      VLLM_API_KEY: ${VLLM_API_KEY:?set VLLM_API_KEY in .env}
      SPEC: ${GPU1_SPEC:-}
      CTX: ${GPU1_CTX:-fast}
      PREFIX_CACHE: ${GPU1_PREFIX_CACHE:-0}
    ports:
      - "18021:18021"
    healthcheck:
      test: ["CMD", "curl", "-sf", "http://127.0.0.1:18021/health"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 900s
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["1"]
              capabilities: [gpu]
    depends_on:
      prepare:
        condition: service_completed_successfully

  # Load balancer. Reaches the replicas over the compose network by service name,
  # so it does not depend on host port publishing. Round-robin, NOT least_conn —
  # see nginx/qwen-lb.conf for why that matters to the A/B.
  lb:
    image: nginx:1.27-alpine
    container_name: qwen38-lb
    restart: unless-stopped
    ports:
      - "${LB_PORT:-18000}:18000"
    volumes:
      - ./nginx/qwen-lb.conf:/etc/nginx/nginx.conf:ro
      - ./logs:/var/log/qwen
    networks: [qwen]
    depends_on:
      - replica-gpu0
      - replica-gpu1
    healthcheck:
      test: ["CMD", "wget", "-qO-", "http://127.0.0.1:18000/lb-health"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 20s

networks:
  qwen:
    name: qwen38-net

volumes:
  # EXTERNAL on purpose: this is the existing cache from the old stack, holding the
  # torch.compile / Triton / FlashInfer JIT artifacts. Reusing it keeps cold starts
  # near 60 s instead of several minutes. Do not let compose create a fresh one.
  qwen-cache:
    external: true
    name: qwen38-rtx3090_qwen-cache