services: strata: image: ghcr.io/niko1221/strata:0.2.9-cuda12 container_name: strata restart: unless-stopped deploy: resources: reservations: devices: - driver: nvidia count: all capabilities: - gpu environment: - TZ=Europe/Moscow - FAMILY=qwen - MODEL=IQ3_XXS - CONTEXT=262144 - KV=k8v4 - VISION=cpu - PARALLEL=1 - TUNE=skip # tune уже отработал и зафиксирован в docker.tune.json — повторно жрёт ~10 мин при каждом пересоздании - PRESET=docker - API_KEY=${STRATA_API_KEY} - STRATA_SET=memory-limit-mib=98304 vram-reserve-mib=2048 expert-cache=auto ports: - 192.168.31.100:8080:8080 shm_size: 2g stop_grace_period: 120s ulimits: memlock: hard: -1 soft: -1 volumes: - /mnt/ssd/apps/strata:/data x-portals: - host: 192.168.31.100 name: Strata path: / port: 8080 scheme: http networks: {}