services:
  llama-server:
    # Pin this. `server-cuda` is a MOVING tag while everything else here is pinned (Pi,
    # adapter, GGUF). The MTP/draft flags below are new and have already changed names once;
    # a silent pull can break startup or drafting determinism. Set LLAMA_IMAGE in .env to an
    # immutable digest before relying on this in production:
    #   docker buildx imagetools inspect ghcr.io/ggml-org/llama.cpp:server-cuda  # copy sha256
    #   LLAMA_IMAGE=ghcr.io/ggml-org/llama.cpp@sha256:<digest>
    image: ${LLAMA_IMAGE:-ghcr.io/ggml-org/llama.cpp:server-cuda}
    container_name: daas-llama
    ports:
      - "8080:8080"
    volumes:
      - ./models:/models
      - ./templates:/templates
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    restart: unless-stopped
    # Qwen3.6-35B-A3B MTP UD-Q4_K_XL on a single L4 24GB. Params follow the reference MTP tune
    # (github.com/hanxiao/Qwen3.6-35B-A3B-MTP-L4): q4_0 KV, --spec-draft-n-max 2 (no p-min),
    # -ub 256 -b 2048, mmap ON + -ngl auto-fit (no --no-mmap, no -ngl 999 -- forcing all layers
    # to GPU OOMs once MTP + KV load). Q4_K_XL weights ~22GB nearly fill the card.
    # This is a HYBRID GDN+MoE model: only 10 of 40 layers carry a per-token KV cache (the rest
    # keep a small fixed recurrent state), so per-token KV is tiny (~0.65GB at q4_0/131072).
    # Real headroom is dominated by the weights, the GDN recurrent-state pool, compute buffers,
    # and the v5-nano embedder (~0.5GB). Measured: ~22.2GB used at full 131072 ctx, no OOM.
    # Lower CTX_SIZE (98304/65536) if nvidia-smi is tight.
    #
    # NOTE: --cache-reuse is intentionally OMITTED from the command below. This hybrid (Gated
    # DeltaNet) model has documented recurrent-state cache drift that can silently corrupt
    # digits across the long --continue + auto-compaction loop (llama.cpp#21681), fatal for a
    # factual dataroom. Re-add `--cache-reuse 256` only once the fix is confirmed in the pinned
    # image and verified by the numeric multi-turn smoke test in docs/DEPLOY.md.
    command: >
      --model /models/${MODEL_FILE:-Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf}
      --host 0.0.0.0
      --port 8080
      --metrics
      --ctx-size ${CTX_SIZE:-131072}
      --parallel 1
      --flash-attn 1
      --cache-type-k q4_0
      --cache-type-v q4_0
      --threads 8
      -ub 256
      -b 2048
      ${SPEC_ARGS:---spec-type draft-mtp --spec-draft-n-max 2}
      --n-predict 8192
      --jinja
      --chat-template-file ${CHAT_TEMPLATE_FILE:-/templates/chat_template.jinja}
      -fitt 512

  daas:
    # Prebuilt image (default): `docker compose pull` fetches it. To build from source instead,
    # `docker compose up -d --build` builds `.` and tags it under this same name.
    image: ${DAAS_IMAGE:-ghcr.io/hanxiao/dataroom:latest}
    build: .
    container_name: daas-app
    ports:
      - "8000:8000"
    env_file:
      - .env
    environment:
      - LLAMA_URL=http://llama-server:8080
      - JOBS_DIR=/data/jobs
      - CONTEXT_WINDOW=${CTX_SIZE:-131072}   # match llama-server --ctx-size (dashboard + compaction)
      - EMBED_DEVICE=cpu           # v5-nano on CPU: frees ~0.5GB so Q4 (~22GB) doesn't OOM the L4
      - TOKENIZERS_PARALLELISM=false
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    volumes:
      - ./data:/data
      # Skills + extensions are read-only methodology. Bind them back over the image copy as
      # :ro so a job (which runs as root in this container) physically cannot create, overwrite,
      # touch or delete anything under pi/ - the kernel returns EROFS. Defense in depth on top of
      # the prompt anchor that already tells the agent the dataroom root is its own job dir.
      - ./pi:/app/pi:ro
    depends_on:
      - llama-server
    restart: unless-stopped

  # Reverse proxy for dataroom.hanxiao.io (Cloudflare -> here). Terminates :80/:443 and
  # forwards to the app. Self-signed origin cert lives in ./nginx/certs (gitignored, made by
  # setup.sh / generate-cert). See nginx/dataroom.conf for the rate-limit + proxy config.
  nginx:
    image: nginx:alpine
    container_name: daas-nginx
    ports:
      - "80:80"
      - "443:443"
    volumes:
      - ./nginx/dataroom.conf:/etc/nginx/conf.d/default.conf:ro
      - ./nginx/certs:/etc/nginx/certs:ro
    depends_on:
      - daas
    restart: unless-stopped