/ods/docker-compose.arc.yml

# ODS — Intel Arc GPU Overlay (oneAPI SYCL, build-from-source)
#
# Builds llama-server from source using Intel oneAPI Base Toolkit.
# Use with: docker compose -f docker-compose.base.yml -f docker-compose.arc.yml up -d
#
# Supported hardware:
#   ARC tier      — Arc A770 16 GB  (and future Arc B-series ≥12 GB)
#   ARC_LITE tier — Arc A750 8 GB, A380 6 GB
#
# Build arguments (customise via .env or shell export):
#   LLAMA_TAG       llama.cpp git tag (default: b8248)
#   ONEAPI_VERSION  oneAPI Base Toolkit image tag (default: 2025.0.0-0-devel-ubuntu22.04)
#   LLAMA_ARC_IMAGE Override to a pre-built image and skip the local build entirely
#                   e.g. LLAMA_ARC_IMAGE=ghcr.io/ggml-org/llama.cpp:server-intel-b8248
#
# First-run note:
#   The SYCL build compiles llama.cpp with Intel icx/icpx. Allow 10–20 min
#   on first `docker compose up --build`. Subsequent starts use Docker cache.
#
# Host prerequisites (Ubuntu/Debian):
#   apt install intel-opencl-icd intel-level-zero-gpu level-zero
#   usermod -aG video,render $USER   # re-login after
#   # Verify GPU: clinfo | grep -i "intel arc"

services:
  llama-server:
    build:
      context: ./images/llama-sycl
      dockerfile: Dockerfile
      args:
        LLAMA_TAG: ${LLAMA_TAG:-b8248}
        ONEAPI_VERSION: ${ONEAPI_VERSION:-2025.0.0-0-devel-ubuntu22.04}
    # Set LLAMA_ARC_IMAGE in .env to skip the local build and pull a pre-built image.
    image: ${LLAMA_ARC_IMAGE:-ods-llama-sycl:local}
    devices:
      - /dev/dri:/dev/dri
    group_add:
      - "${VIDEO_GID:-44}"
      - "${RENDER_GID:-992}"
    environment:
      # Level Zero selects the Intel Arc GPU; persistent cache avoids
      # recompiling SYCL kernels on every container start (~30 s savings).
      - ONEAPI_DEVICE_SELECTOR=level_zero:gpu
      - SYCL_CACHE_PERSISTENT=1
      # Enable Intel GPU System Management Interface for telemetry
      - ZES_ENABLE_SYSMAN=1
    command:
      - --model
      - /models/${GGUF_FILE:-Qwen3.5-9B-Q4_K_M.gguf}
      - --host
      - 0.0.0.0
      - --port
      - "8080"
      - --n-gpu-layers
      - "${N_GPU_LAYERS:-99}"
      - --ctx-size
      - "${CTX_SIZE:-32768}"
      - --metrics
    deploy:
      resources:
        limits:
          cpus: '${LLAMA_CPU_LIMIT:-16.0}'
          memory: ${LLAMA_SERVER_MEMORY_LIMIT:-24G}
        reservations:
          cpus: '${LLAMA_CPU_RESERVATION:-2.0}'
          memory: 4G

  dashboard-api:
    environment:
      # Hard-code sycl so the dashboard uses Intel sysfs GPU detection
      # regardless of .env state.
      - GPU_BACKEND=sycl
    volumes:
      - /sys/class/drm:/sys/class/drm:ro
      - /sys/class/hwmon:/sys/class/hwmon:ro