# LiteLLM proxy config for DeepSeek V4.1 Flash on a vLLM server.
# Replace SPARK_HOST with your server's address. Model names are what clients see.
model_list:
  - model_name: deepseek-v4.1-flash
    litellm_params:
      # hosted_vllm keeps reasoning_content on assistant messages in v1.99.x;
      # v1.100.0 through v1.104.0 strip it (BerriAI/litellm#41392).
      model: hosted_vllm/deepseek-v4.1-flash
      api_base: http://SPARK_HOST:8000/v1
      api_key: "none"
      temperature: 1.0   # model card
      top_p: 0.95        # model card
      extra_body:
        chat_template_kwargs:
          reasoning_effort: 75   # DeepSeek's reference default; integer 1-100

litellm_settings:
  callbacks: loopguard.proxy_handler_instance
  drop_params: false

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY