Support default chat template kwargs

lightllm/server/api_cli.py

@@ -1,4 +1,15 @@
 import argparse
+import json
+
+
+def _json_dict(value: str) -> dict:
+    try:
+        parsed = json.loads(value)
+    except json.JSONDecodeError as e:
+        raise argparse.ArgumentTypeError(f"invalid JSON object: {e}") from e
+    if not isinstance(parsed, dict):
+        raise argparse.ArgumentTypeError("value must be a JSON object")
+    return parsed
 
 
 def make_argument_parser() -> argparse.ArgumentParser:
@@ -202,6 +213,16 @@ def make_argument_parser() -> argparse.ArgumentParser:
             "- /test/chat_template/tool_chat_template_deepseekr1.jinja"
         ),
     )
+    parser.add_argument(
+        "--default_chat_template_kwargs",
+        "--default-chat-template-kwargs",
+        type=_json_dict,
+        default=None,
+        help=(
+            "Default JSON kwargs passed to tokenizer.apply_chat_template for every chat request. "
+            'For example: \'{"preserve_thinking": true}\'. Request chat_template_kwargs override these values.'
+        ),
+    )
 
     parser.add_argument(
         "--running_max_req_size", type=int, default=256, help="the max size for forward requests in the same time"

lightllm/server/api_openai.py

@@ -12,7 +12,7 @@
 from lightllm.server.reasoning_parser import ReasoningParser
 
 from .function_call_parser import TOOLS_TAG_LIST, FunctionCallParser, ToolCallItem
-from .build_prompt import build_prompt, init_tokenizer
+from .build_prompt import build_prompt, get_effective_chat_template_kwargs, init_tokenizer
 
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 import ujson as json
@@ -163,13 +163,14 @@ def _is_force_thinking_mode(request: ChatCompletionRequest) -> bool:
     reasoning_parser = get_env_start_args().reasoning_parser
     if not reasoning_parser:
         return False
+    chat_template_kwargs = get_effective_chat_template_kwargs(request)
     if reasoning_parser in ["qwen3-thinking", "gpt-oss", "minimax"]:
         return True
     if reasoning_parser in ["deepseek-v3"]:
-        return request.chat_template_kwargs is not None and request.chat_template_kwargs.get("thinking") is True
+        return chat_template_kwargs.get("thinking") is True
     if reasoning_parser in ["qwen3", "glm45", "nano_v3", "interns1", "gemma4"]:
         # qwen3, glm45, nano_v3, interns1, and gemma4 are reasoning by default;
-        return not request.chat_template_kwargs or request.chat_template_kwargs.get("enable_thinking", True) is True
+        return chat_template_kwargs.get("enable_thinking", True) is True
     return True  # default
 
 

lightllm/server/build_prompt.py

@@ -4,6 +4,7 @@
 from lightllm.utils.log_utils import init_logger
 from functools import lru_cache
 from lightllm.utils.config_utils import get_model_type_v1
+from lightllm.utils.envs_utils import get_env_start_args
 
 logger = init_logger(__name__)
 
@@ -126,6 +127,16 @@ def _normalize_multimodal_content_types(messages: list) -> None:
                 part["type"] = "audio"
 
 
+def get_effective_chat_template_kwargs(request) -> dict:
+    kwargs = {}
+    default_kwargs = getattr(get_env_start_args(), "default_chat_template_kwargs", None)
+    if default_kwargs:
+        kwargs.update(default_kwargs)
+    if request.chat_template_kwargs:
+        kwargs.update(request.chat_template_kwargs)
+    return kwargs
+
+
 async def build_prompt(request, tools) -> str:
     # pydantic格式转成dict， 否则，当根据tokenizer_config.json拼template时，Jinja判断无法识别
     messages = [m.model_dump(by_alias=True, exclude_none=True) for m in request.messages]
@@ -141,8 +152,9 @@ async def build_prompt(request, tools) -> str:
     if request.role_settings:
         kwargs["role_setting"] = request.role_settings
 
-    if request.chat_template_kwargs:
-        kwargs.update(request.chat_template_kwargs)
+    chat_template_kwargs = get_effective_chat_template_kwargs(request)
+    if chat_template_kwargs:
+        kwargs.update(chat_template_kwargs)
 
     # 修复一些parser类型是默认打开thinking，但是 tokenizer有时候不知道打开了thinking。导致
     # 构建的reasoning parser 和 tokenizer 的行为不对齐导致的问题。

lightllm/server/core/objs/start_args_type.py

@@ -1,5 +1,5 @@
 from dataclasses import dataclass, field
-from typing import List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple
 
 # 只是为了更好的编程提示
 
@@ -57,6 +57,7 @@ class StartArgs:
         },
     )
     chat_template: Optional[str] = field(default=None)
+    default_chat_template_kwargs: Optional[Dict] = field(default=None)
     running_max_req_size: int = field(default=512)
     tp: int = field(default=1)
     dp: int = field(default=1)