dexogen/luce-dflash-server repository overview

services:
  luce-qwen27:
    image: dexogen/luce-dflash-server:latest
    container_name: luce-qwen27
    restart: unless-stopped
    ports:
      - "8081:8080"
    volumes:
      - ./models:/models:ro
      - ./src/lucebox-hub:/src:ro
      - ./hf-cache:/root/.cache/huggingface
    environment:
      MODEL_NAME: qwen3.6:27b
      MODEL_CONTEXT_LENGTH: "131072"
      NVIDIA_VISIBLE_DEVICES: "GPU-82371e50-134a-7a7b-3b30-bfdcf4a8622b"
      NVIDIA_DRIVER_CAPABILITIES: "compute,utility"
      LD_LIBRARY_PATH: /src/dflash/build/deps/llama.cpp/ggml/src:/src/dflash/build/deps/llama.cpp/ggml/src/ggml-cuda:/src/dflash/build/deps/llama.cpp/src
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids:
                - "GPU-82371e50-134a-7a7b-3b30-bfdcf4a8622b"
              capabilities: [gpu]
    working_dir: /src/dflash
    command:
      - python3
      - scripts/server.py
      - --host
      - 0.0.0.0
      - --port
      - "8080"
      - --target
      - /models/luce-qwen27/Qwen3.6-27B-Q4_K_M.gguf
      - --draft
      - /models/luce-qwen27/draft/model.safetensors
      - --bin
      - /src/dflash/build/test_dflash
      - --budget
      - "16"
      - --max-ctx
      - "16384"
      - --tokenizer
      - Qwen/Qwen3.6-27B
      - --prefill-compression
      - auto
      - --prefill-threshold
      - "32000"
      - --prefill-keep-ratio
      - "0.10"
      - --prefill-drafter
      - /models/luce-qwen27/pflash/Qwen3-0.6B-BF16.gguf
      - --prefill-drafter-tokenizer
      - Qwen/Qwen3-0.6B