Ollama Rocm Dokploy

docker-compose.yml

version: "3.8"

services:
  ollama:
    image: ollama/ollama:rocm
    restart: unless-stopped
    privileged: true
    dns:
      - 8.8.8.8
      - 1.1.1.1
    deploy:
      resources:
        limits:
          cpus: "14" # Leaves 1 core for the host system
          memory: 24g # Keep high for APU shared memory
    devices:
      - /dev/kfd:/dev/kfd
      - /dev/dri:/dev/dri
    group_add:
      - video
    shm_size: "8gb"
    volumes:
      - ollama:/root/.ollama
    ports:
      - "11434:11434"
    environment:
      # -----------------------------------------------------
      # APU STABILITY
      # -----------------------------------------------------
      - LLAMA_HIP_UMA=1
      - HSA_OVERRIDE_GFX_VERSION=11.0.0
      - HSA_ENABLE_SDMA=0
      - GPU_MAX_HW_QUEUES=2

      # -----------------------------------------------------
      # SPEED OPTIMIZATIONS (For Visible Reasoning)
      # -----------------------------------------------------
      - OLLAMA_FLASH_ATTENTION=1     # CRITICAL: Makes long reasoning chains generate faster
      - OLLAMA_KV_CACHE_TYPE=q8_0    # OPTIMIZED: Uses less memory bandwidth than f16, keeping tokens/s high
      - OLLAMA_NUM_PARALLEL=4        # INCREASED: Allows multiple requests to utilize your 15 CPU cores
      - OLLAMA_MAX_LOADED_MODELS=1   # Prevents swapping
    networks:
      - dokploy-network

volumes:
  ollama:

networks:
  dokploy-network:
    external: true