From 5a38f4a2fa0a38e6d4fe52cbec135686df860af9 Mon Sep 17 00:00:00 2001
From: root
 <root@cc-adc62d1c-476c-4ee4-9647-0c011c0b6d70-0.cc-adc62d1c-476c-4ee4-9647-0c011c0b6d70.ws-19969783038.svc.cluster.local>
Date: Wed, 2 Sep 2026 09:08:38 +0000
Subject: [PATCH] baseline8 fix d03

---
 qwen3_6_scripts/serving_chat.py | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/qwen3_6_scripts/serving_chat.py b/qwen3_6_scripts/serving_chat.py
index b26b58f5..e8a3d9aa 100644
--- a/qwen3_6_scripts/serving_chat.py
+++ b/qwen3_6_scripts/serving_chat.py
@@ -316,6 +316,16 @@ class OpenAIServingChat(OpenAIServing):
                 tool.model_dump() for tool in request.tools
             ]
 
+            # [BI100] Ensure enable_thinking defaults to True when the
+            # request does not specify it.  The Qwen3.5 chat template
+            # may disable thinking when tools are present and the flag
+            # is absent, which prevents the model from reasoning about
+            # tool selection and generating <tool_call> tags.
+            if request.chat_template_kwargs is None:
+                request.chat_template_kwargs = {}
+            if "enable_thinking" not in request.chat_template_kwargs:
+                request.chat_template_kwargs["enable_thinking"] = True
+
             prompt: Union[str, List[int]]
             is_mistral_tokenizer = isinstance(tokenizer, MistralTokenizer)
             if is_mistral_tokenizer: