x

README.md

@@ -30,6 +30,7 @@
 
 - `services.openai.host` / `services.openai.port`：OpenAI 协议服务监听地址与端口（默认 `0.0.0.0:8001`）
 - `public_model_name`：对外固定模型名，切换底层模型时可保持调用方参数不变
+- `default_enable_thinking`：服务端默认思考开关，默认 `false`（即调用方不传时也关闭）
 - `reasoning_enabled`：是否启用推理解析器参数注入，默认 `false`
 - `api_key`：OpenAI 接口访问密钥
 - `tensor_parallel_size`：张量并行数，双卡建议 `2`
@@ -93,6 +94,7 @@ curl -X POST "http://localhost:8001/v1/chat/completions" \
 - API Key 使用 `config.json` 中 `api_key`
 - 模型名固定使用 `config.json` 中 `public_model_name`（默认 `Qwen_local_model`）
 - 思考模式按请求控制：`chat_template_kwargs.enable_thinking=false/true`
+- 若调用方未传 `chat_template_kwargs.enable_thinking`，服务端使用 `default_enable_thinking` 兜底
 - 仅当模型需要推理解析器时，再将 `config.json` 中 `reasoning_enabled` 设为 `true`
 - 若使用工具调用，`config.json` 中应配置 `tool_call_parser` 与 `enable_auto_tool_choice`
 - 服务强制离线模式，不会回退到 Hugging Face 远程下载

app/config.py

@@ -19,6 +19,7 @@ class Settings(BaseModel):
     openai_port: int = 8001
     vllm_openai_internal_url: str = "http://127.0.0.1:8001/v1"
     public_model_name: str = "Qwen_local_model"
+    default_enable_thinking: bool = False
     reasoning_enabled: bool = False
     model_root: str = "/opt/model"
     offline_mode: bool = True
@@ -49,6 +50,7 @@ def get_settings() -> Settings:
         openai_port=runtime["openai_port"],
         vllm_openai_internal_url=runtime["vllm_openai_internal_url"],
         public_model_name=runtime["public_model_name"],
+        default_enable_thinking=runtime["default_enable_thinking"],
         reasoning_enabled=runtime["reasoning_enabled"],
         model_root=runtime["model_root"],
         offline_mode=runtime["offline_mode"],

app/model_catalog.py

@@ -74,6 +74,7 @@ def resolve_runtime_settings(content: dict[str, Any]) -> dict[str, Any]:
         "openai_port": openai_port,
         "vllm_openai_internal_url": internal_url.rstrip("/"),
         "public_model_name": _to_str(content.get("public_model_name"), "Qwen_local_model"),
+        "default_enable_thinking": _to_bool(content.get("default_enable_thinking"), False),
         "api_key": str(content.get("api_key", "")).strip() or None,
         "reasoning_enabled": _to_bool(content.get("reasoning_enabled"), False),
         "tensor_parallel_size": _to_int(content.get("tensor_parallel_size"), 2),

app/start_openai.py

@@ -1,3 +1,4 @@
+import json
 import os
 import subprocess
 import sys
@@ -19,6 +20,7 @@ def build_command() -> list[str]:
     host = str(runtime["openai_host"])
     port = str(runtime["openai_port"])
     public_model_name = str(runtime["public_model_name"]).strip()
+    default_enable_thinking = bool(runtime["default_enable_thinking"])
     reasoning_enabled = bool(runtime["reasoning_enabled"])
     api_key = runtime["api_key"] or ""
     dtype = str(updates["dtype"] or runtime["dtype"])
@@ -56,6 +58,12 @@ def build_command() -> list[str]:
         cmd.append("--enable-auto-tool-choice")
     if updates["tool_call_parser"]:
         cmd.extend(["--tool-call-parser", str(updates["tool_call_parser"])])
+    cmd.extend(
+        [
+            "--default-chat-template-kwargs",
+            json.dumps({"enable_thinking": default_enable_thinking}),
+        ]
+    )
     if reasoning_enabled and reasoning_parser:
         cmd.extend(["--reasoning-parser", reasoning_parser])
     if quantization:

config.json

@@ -10,6 +10,7 @@
     }
   },
   "public_model_name": "Qwen_local_model",
+  "default_enable_thinking": false,
   "reasoning_enabled": false,
   "api_key": "sk-szcjw",
   "tensor_parallel_size": 2,