# LiteLLM Proxy · 统一 OpenAI 兼容网关
# 上游：vLLM(qwen3.6-27b) + Xinference(bge-m3)

model_list:
  # ---- 对话 / 补全（183 vLLM）----
  - model_name: qwen3.6-27b
    litellm_params:
      model: openai/qwen3.6-27b
      api_base: http://116.62.10.183:8000/v1
      api_key: "123456"
      timeout: 300
      max_retries: 2
      # Qwen3 默认会产出 thinking；网关侧默认关掉，避免占满 max_tokens
      extra_body:
        chat_template_kwargs:
          enable_thinking: false

  # 别名，方便客户端用常见名字
  - model_name: qwen
    litellm_params:
      model: openai/qwen3.6-27b
      api_base: http://116.62.10.183:8000/v1
      api_key: "123456"
      timeout: 300
      max_retries: 2
      extra_body:
        chat_template_kwargs:
          enable_thinking: false

  # ---- 学习用新增示例：对外新名字，仍转发到同一上游 ----
  # model_name = 客户端调用时填的名字（LiteLLM 对外暴露）
  # litellm_params.model = openai/<上游真实模型 id>
  # api_base / api_key = 上游地址与密钥
  - model_name: demo-chat
    litellm_params:
      model: openai/qwen3.6-27b
      api_base: http://116.62.10.183:8000/v1
      api_key: "123456"
      timeout: 300
      max_retries: 2
      extra_body:
        chat_template_kwargs:
          enable_thinking: false

  # ---- 向量（207 Xinference）----
  - model_name: bge-m3
    litellm_params:
      model: openai/bge-m3
      api_base: http://host.docker.internal:9997/v1
      api_key: "xf-K8H3LiS27H4188XdOyWJj74MWMOlIR9tZbiybDbxbSas7eAI"
      timeout: 120

  # ---- 学习用：负载均衡 + 手动指定机器 ----
  # api_base 故意指向不存在端口，仅模拟；真实调用失败属预期
  #
  # ① 自动负载均衡：同名两条 → 客户端 model=demo-lb，由 router 选机
  - model_name: demo-lb
    litellm_params:
      model: openai/fake-qwen-a
      api_base: http://127.0.0.1:18001/v1
      api_key: "fake-key-machine-a"
      timeout: 10
      max_retries: 0

  - model_name: demo-lb
    litellm_params:
      model: openai/fake-qwen-b
      api_base: http://127.0.0.1:18002/v1
      api_key: "fake-key-machine-b"
      timeout: 10
      max_retries: 0

  # ② 手动指定：唯一 model_name 钉死某一台（排障 / 灰度 / 压测常用）
  #    同一台机器可既挂在 demo-lb 里参与均衡，又单独起一个别名
  - model_name: demo-lb-a
    litellm_params:
      model: openai/fake-qwen-a
      api_base: http://127.0.0.1:18001/v1
      api_key: "fake-key-machine-a"
      timeout: 10
      max_retries: 0

  - model_name: demo-lb-b
    litellm_params:
      model: openai/fake-qwen-b
      api_base: http://127.0.0.1:18002/v1
      api_key: "fake-key-machine-b"
      timeout: 10
      max_retries: 0

litellm_settings:
  drop_params: true
  set_verbose: false
  request_timeout: 300
  num_retries: 2

# 同名多部署时的分流策略（simple-shuffle = 随机选一台）
router_settings:
  routing_strategy: simple-shuffle

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY