【專案需求】

一般使用者： 透過 Librechat 進行日常問答。

RD 工程師： 透過 API 串接模型，輔助 Code 編寫。

【目前遇到的問題】
目前我們部署了 Qwen3.5:9B 以及 Qwen3.6:27B 兩個模型，雖然已經嘗試開啟了 fp8 Cache，但實際測試下來，兩個模型的推論生成速度皆不盡理想，無法滿足 RD 開發時需要的流暢度。

想請教各位前輩在不購買設備的情況下還有甚麼優化空間呢?

services:
  vllm-qwen-27b:
    image: vllm/vllm-openai:latest
    container_name: vllm-qwen-27b
    restart: unless-stopped
    ipc: host
    shm_size: "16gb"
    ports:
      - "8001:8000"
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    environment:
      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN}
    command: >
      Qwen/Qwen3.6-27B
      --host 0.0.0.0
      --port 8000
      --served-model-name qwen3.6-27b
      --gpu-memory-utilization 0.70
      --max-model-len 32768
      --kv-cache-dtype fp8
      --trust-remote-code
      --reasoning-parser qwen3
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  vllm-qwen:
    image: vllm/vllm-openai:qwen3_5-cu130
    container_name: vllm-qwen
    restart: unless-stopped
    ipc: host
    shm_size: "8gb"
    ports:
      - "8002:8000"
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    environment:
      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN}
    command: >
      Qwen/Qwen3.5-9B
      --host 0.0.0.0
      --port 8000
      --served-model-name qwen3.5-9b
      --gpu-memory-utilization 0.25
      --max-model-len 16384
      --kv-cache-dtype fp8
      --trust-remote-code
      --reasoning-parser qwen3
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # ============================================================
  # LibreChat：主要聊天介面
  # ============================================================
  librechat:
    image: ghcr.io/danny-avila/librechat:latest
    container_name: librechat
    restart: unless-stopped
    ports:
      - "3000:3080"
    volumes:
      - ./librechat-config/librechat.yaml:/app/librechat.yaml
      - librechat-uploads:/app/client/public/images
      - librechat-logs:/app/api/logs
    environment:
      - MONGO_URI=mongodb://mongodb:27017/LibreChat
      - MEILI_HOST=http://meilisearch:7700
      - MEILI_MASTER_KEY=${MEILI_MASTER_KEY}
      - JWT_SECRET=${JWT_SECRET}
      - JWT_REFRESH_SECRET=${JWT_REFRESH_SECRET}
      - ALLOW_REGISTRATION=true       # 首次設定完可改 false
      - RAG_USE_FULL_CONTEXT=true     # ← 關鍵：強制全文送入，不切塊
    depends_on:
      - vllm-qwen-27b
      - vllm-qwen

  # LibreChat 需要的 MongoDB
  mongodb:
    image: mongo:6
    container_name: librechat-mongo
    restart: unless-stopped
    volumes:
      - mongodb-data:/data/db

  # LibreChat 需要的 MeiliSearch（對話搜尋功能）
  meilisearch:
    image: getmeili/meilisearch:v1.12
    container_name: librechat-meili
    restart: unless-stopped
    environment:
      - MEILI_MASTER_KEY=${MEILI_MASTER_KEY}
      - MEILI_NO_ANALYTICS=true
    volumes:
      - meilisearch-data:/meili_data

volumes:
  librechat-uploads:
  librechat-logs:
  mongodb-data:
  meilisearch-data: