From 6c514ab0257e9d153b5795bfa341f64c89584d32 Mon Sep 17 00:00:00 2001
From: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Date: Tue, 15 Sep 2026 13:16:26 +0800
Subject: [PATCH] Force reasoning mode for GLM-5.3 chat templates (#39227)

Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
---
 python/sglang/srt/parser/template_detection.py       | 7 +++++++
 test/registered/unit/parser/test_template_manager.py | 7 +++++++
 2 files changed, 14 insertions(+)

diff --git a/python/sglang/srt/parser/template_detection.py b/python/sglang/srt/parser/template_detection.py
index 9dda09563..a662492f0 100644
--- a/python/sglang/srt/parser/template_detection.py
+++ b/python/sglang/srt/parser/template_detection.py
@@ -168,6 +168,13 @@ REASONING_MODE_RULES = (
             and not ctx.has_text("thinking")
         ),
     ),
+    DetectionRule(
+        name="glm53_always_think",
+        value=ReasoningToggleConfig(special_case="always"),
+        # GLM-5.3's generation prompt opens <think>, so the output carries no
+        # opening tag; lambda because _is_glm53 is defined below.
+        predicate=lambda ctx: _is_glm53(ctx),
+    ),
     DetectionRule(
         name="mistral_reasoning_effort",
         value=ReasoningToggleConfig(special_case="mistral"),
diff --git a/test/registered/unit/parser/test_template_manager.py b/test/registered/unit/parser/test_template_manager.py
index 39d844e41..1d1da64fa 100644
--- a/test/registered/unit/parser/test_template_manager.py
+++ b/test/registered/unit/parser/test_template_manager.py
@@ -116,6 +116,13 @@ class TestTemplateManagerReasoningDetection(unittest.TestCase):
                     "glm47",
                 )
 
+    def test_glm53_effort_template_forces_reasoning(self):
+        vocab = ["<tool_call>", "<arg_key>", "<arg_value>", "<|user|>", "<|endoftext|>"]
+        force, config, _ = self._detect(_glm53_template("+"), vocab)
+
+        self.assertTrue(force)
+        self.assertEqual(config, ReasoningToggleConfig(special_case="always"))
+
     def test_interns1_detects_enable_thinking_default_true(self):
         template = """
         {% set default_thinking_sys %}...<think>...</think>{% endset %}