# model_config.yaml - 多模型配置示例
# 默认启动的模型
default: "deepseek_r1_distill_qwen_32b_awq"

# 模型配置
models:
  deepseek_r1_distill_qwen_14b:
    path: "/app/vllm/models/DeepSeek-R1-Distill-Qwen-14B"
    name: "DeepSeek-R1-Distill-Qwen-14B"
    max_model_len: 8192
    gpu_memory_utilization: 0.9
    port: 2001
    dtype: "float16"
    quantization: "awq"
    tensor_parallel_size: 1
    enforce_eager: true
    api_key: "sk-14b-20240101-abcdef123456"
    
  deepseek_r1_distill_qwen_32b_awq:
    path: "/app/vllm/models/DeepSeek-R1-Distill-Qwen-32B-AWQ"
    name: "DeepSeek-R1-Distill-Qwen-32B-AWQ"
    
    # 模型性能参数
    max_model_len: 32768
    gpu_memory_utilization: 0.95
    enforce_eager: true
    max_num_seqs: 2
    max_num_batched_tokens: 1024
    block_size: 16
    tensor_parallel_size: 1
    swap_space: 0
    
    # 新增：采样参数默认值
    sampling_defaults:
      temperature: 0.6
      max_tokens: 4096
      top_p: 0.9
      frequency_penalty: 0.0
      presence_penalty: 0.0
      stop:
        - "用户:"
        - "助手:"
        - "###"
        - "问题:"
        - "回答:"
    
    # 其他配置
    dtype: "auto"
    quantization: "awq"
    port: 2001
    api_key: "sk-32b-20240101-ghijk789012"
    
  glm_4_7_flash_awq:
    path: "/app/vllm/models/GLM-4.7-Flash-AWQ"
    name: "GLM-4.7-Flash-AWQ"

    # 模型性能参数
    max_model_len: 32768
    gpu_memory_utilization: 0.9
    enforce_eager: false
    max_num_seqs: 2
    max_num_batched_tokens: 1024
    block_size: 16
    tensor_parallel_size: 1
    swap_space: 0
    
    # 新增：采样参数默认值
    sampling_defaults:
      temperature: 0.6
      max_tokens: 4096
      top_p: 0.9
      frequency_penalty: 0.0
      presence_penalty: 0.0
      stop:
        - "用户:"
        - "助手:"
        - "###"
        - "问题:"
        - "回答:"
    
    # 其他配置
    dtype: "auto"
    quantization: "awq"
    port: 2001
    api_key: "sk-32b-20240101-ghijk789012"

  qwen3_vl_32b_instruct_awq:
    path: "/app/vllm/models/Qwen3-VL-32B-Instruct-AWQ"
    name: "Qwen3-VL-32B-Instruct-AWQ"
    max_model_len: 32768
    gpu_memory_utilization: 0.7
    port: 2001
    dtype: "auto"
    quantization: "awq"
    tensor_parallel_size: 1
    enforce_eager: true
    api_key: "sk-glm-20240101-lmnop345678"

# 服务器通用设置
server:
  host: "0.0.0.0"
  log_level: "info"
  # 全局管理员密钥（拥有所有模型的访问权限）
  admin_key: "sk-admin-20240101-xyz789"
  # 允许的请求头名称（支持多个，按顺序检查）
  api_key_headers: ["Authorization", "X-API-Key", "api-key"]
  # 是否允许通过查询参数传递密钥
  allow_query_param: true
  # 查询参数名称
  api_key_param: "api_key"