POST /v1/chat/completions   （OpenAI 兼容；多轮上下文请把历史一起传）
{
  "model": "deepseek-r1-distill-qwen-32b",
  "messages": [
    {"role":"system","content":"你是一个简洁的中文助手"},
    {"role":"user","content":"你好"},
    {"role":"assistant","content":"你好！"},
    {"role":"user","content":"我刚才说了什么？"}
  ],
  "stream": true,
  "max_tokens": 4096
}

流式（SSE）：思考与正文分开回传
  data: {"choices":[{"delta":{"reasoning_content":"思考中…"}}]}
  data: {"choices":[{"delta":{"content":"答案…"}}]}
  data: {"done":true}
  data: [DONE]

非流式（stream:false）：
  {"choices":[{"message":{"content":"…","reasoning_content":"…"}}]}

简版接口 POST /api/chat 同参，SSE 为 {"reasoning":"…"} / {"content":"…"}