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
+- `reasoning_enabled`：思考推理开关，默认 `false`（关闭）
 - `api_key`：OpenAI 接口访问密钥
 - `tensor_parallel_size`：张量并行数，双卡建议 `2`
 - `dtype`：推理精度，默认 `bfloat16`
@@ -44,7 +45,7 @@
 - `models.default`：默认模型名
 - `models.selected`：当前生效模型名
 - `models.profiles`：模型配置集合
-- 每个模型必须包含：`local_path`，并建议补充 `ctx`、`max_num_seqs`、`max_tokens`、`dtype`、`quantization`
+- 每个模型必须包含：`local_path`，并建议补充 `ctx`、`max_num_seqs`、`max_tokens`、`dtype`、`quantization`、`reasoning_parser`
 
 启动时会按以下优先级选模型：
 
@@ -91,6 +92,7 @@ curl -X POST "http://localhost:8001/v1/chat/completions" \
 - Base URL 使用 `http://<服务器IP>:8001/v1`
 - 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`）
+- 若要启用思考推理，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
 - 所有路径按 Ubuntu 规范填写，本地模型建议使用 `/opt/model/<模型目录>`

app/config.py

@@ -18,6 +18,7 @@ class Settings(BaseModel):
     openai_host: str = "0.0.0.0"
     openai_port: int = 8001
     public_model_name: str = "Qwen_local_model"
+    reasoning_enabled: bool = False
     model_root: str = "/opt/model"
     offline_mode: bool = True
     max_model_len: int = 8192
@@ -46,6 +47,7 @@ def get_settings() -> Settings:
         openai_host=runtime["openai_host"],
         openai_port=runtime["openai_port"],
         public_model_name=runtime["public_model_name"],
+        reasoning_enabled=runtime["reasoning_enabled"],
         model_root=runtime["model_root"],
         offline_mode=runtime["offline_mode"],
         api_key=runtime["api_key"],

app/model_catalog.py

@@ -70,6 +70,7 @@ def resolve_runtime_settings(content: dict[str, Any]) -> dict[str, Any]:
         "openai_port": _to_int(openai_service.get("port"), 8001),
         "public_model_name": _to_str(content.get("public_model_name"), "Qwen_local_model"),
         "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),
         "dtype": str(content.get("dtype", "bfloat16")),
         "revision": str(content.get("revision", "")).strip() or None,
@@ -103,6 +104,7 @@ def resolve_model_profile(
         "served_model_name": profile.get("served_model_name", model_key),
         "dtype": _to_str(profile.get("dtype")),
         "quantization": _to_str(profile.get("quantization")),
+        "reasoning_parser": _to_str(profile.get("reasoning_parser")),
         "max_model_len": _to_int(profile.get("ctx"), 8192),
         "max_num_seqs": _to_int(profile.get("max_num_seqs"), 64),
         "max_tokens": _to_int(profile.get("max_tokens"), 4096),

app/start_openai.py

@@ -19,9 +19,11 @@ def build_command() -> list[str]:
     host = str(runtime["openai_host"])
     port = str(runtime["openai_port"])
     public_model_name = str(runtime["public_model_name"]).strip()
+    reasoning_enabled = bool(runtime["reasoning_enabled"])
     api_key = runtime["api_key"] or ""
     dtype = str(updates["dtype"] or runtime["dtype"])
     quantization = str(updates["quantization"] or "").strip()
+    reasoning_parser = str(updates["reasoning_parser"] or "").strip()
     revision = runtime["revision"] or ""
     cmd = [
         sys.executable,
@@ -54,6 +56,8 @@ 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"])])
+    if reasoning_enabled and reasoning_parser:
+        cmd.extend(["--reasoning-parser", reasoning_parser])
     if quantization:
         cmd.extend(["--quantization", quantization])
     if revision:

config.json

@@ -10,6 +10,7 @@
     }
   },
   "public_model_name": "Qwen_local_model",
+  "reasoning_enabled": false,
   "api_key": "sk-szcjw",
   "tensor_parallel_size": 2,
   "dtype": "bfloat16",
@@ -33,6 +34,7 @@
           "VLLM_USE_TRITON_AWQ": "1"
         },
         "tool_call_parser": "qwen3_xml",
+        "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
         "served_model_name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
         "hf_model_id": "cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit"
@@ -46,6 +48,7 @@
         "max_tokens": "32768",
         "gpu_util": "0.98",
         "tool_call_parser": "qwen3_xml",
+        "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
         "served_model_name": "GLM-4.7-Flash-AWQ",
         "hf_model_id": "THUDM/GLM-4.7-Flash-AWQ"
@@ -60,6 +63,7 @@
         "max_tokens": "8192",
         "gpu_util": "0.94",
         "tool_call_parser": "qwen3_xml",
+        "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
         "served_model_name": "Qwen3.5-27B-FP8",
         "hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
@@ -70,12 +74,13 @@
         "quantization": "gptq",
         "ctx": "65536",
         "trust_remote": true,
-        "valid_tp": [1, 2],
+        "valid_tp": [2],
         "max_num_seqs": "2",
-        "max_tokens": "8192",
+        "max_tokens": "2048",
         "gpu_util": "0.90",
         "enforce_eager": false,
         "tool_call_parser": "qwen3_xml",
+        "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
         "served_model_name": "Qwen3.5-35B-A3B-GPTQ-Int4",
         "hf_model_id": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4"