Auto-detect GLM-5.3 chat templates as glm45/glm47 parsers (#38297)

python/sglang/srt/parser/template_detection.py

@@ -336,9 +336,26 @@ def _is_glm45(ctx):
     )
 
 
+def _is_glm53(ctx):
+    # GLM-5.3 keeps the GLM-4.5 prompt and tool-call format but replaces the
+    # enable_thinking toggle with an always-on "Reasoning Effort:" header.
+    return (
+        ctx.has_text("[gMASK]<sop>")
+        and ctx.has_text("Reasoning Effort:")
+        and not ctx.has_text("enable_thinking")
+        and ctx.has_text("<tool_call>")
+        and ctx.has_text("<arg_key>")
+        and ctx.has_text("<arg_value>")
+    )
+
+
+def _is_glm_family(ctx):
+    return _is_glm45(ctx) or _is_glm53(ctx)
+
+
 def _is_glm47(ctx):
-    return _is_glm45(ctx) and ctx.has_pattern(
-        r"\{\{[-\s]*['\"]<tool_call>['\"]\s*\+\s*tc\.name"
+    return _is_glm_family(ctx) and ctx.has_pattern(
+        r"\{\{[-\s]*['\"]<tool_call>['\"]\s*[+~]\s*tc\.name"
     )
 
 
@@ -486,7 +503,7 @@ REASONING_PARSER_RULES = (
     DetectionRule(name="gpt_oss", value="gpt-oss", predicate=_is_gpt_oss),
     DetectionRule(name="kimi_k2", value="kimi_k2", predicate=_is_kimi_k2),
     DetectionRule(name="nemotron_3", value="nemotron_3", predicate=_is_nemotron_3),
-    DetectionRule(name="glm45", value="glm45", predicate=_is_glm45),
+    DetectionRule(name="glm45", value="glm45", predicate=_is_glm_family),
     DetectionRule(name="hunyuan", value="hunyuan", predicate=_is_hunyuan),
     DetectionRule(name="poolside_v1", value="poolside_v1", predicate=_is_poolside_v1),
     DetectionRule(name="mimo", value="mimo", predicate=_is_mimo),
@@ -526,7 +543,7 @@ TOOL_CALL_PARSER_RULES = (
     DetectionRule(name="deepseek_v31", value="deepseekv31", predicate=_is_deepseek_v31),
     DetectionRule(name="lfm2", value="lfm2", predicate=_is_lfm2),
     DetectionRule(name="glm47", value="glm47", predicate=_is_glm47),
-    DetectionRule(name="glm45", value="glm45", predicate=_is_glm45),
+    DetectionRule(name="glm45", value="glm45", predicate=_is_glm_family),
     DetectionRule(name="minicpm5", value="minicpm5", predicate=_is_minicpm5),
     DetectionRule(name="hunyuan", value="hunyuan", predicate=_is_hunyuan),
     DetectionRule(name="poolside_v1", value="poolside_v1", predicate=_is_poolside_v1),

test/registered/unit/parser/test_template_manager.py

@@ -38,6 +38,26 @@ def _patch_hf_transformers_utils(get_tokenizer, get_config=None):
     return patch.dict(sys.modules, {module.__name__: module})
 
 
+def _glm53_template(concat):
+    """GLM-5.3 shape: always-on thinking behind a ``Reasoning Effort:`` header
+    (no ``enable_thinking`` toggle) and the compact GLM-4.7 tool-call format;
+    HF revisions differ only in the ``+`` / ``~`` concat operator."""
+    return (
+        "[gMASK]<sop>\n"
+        "{%- set effective_reasoning_effort = reasoning_effort if reasoning_effort is defined "
+        "and reasoning_effort in ['low', 'high'] else 'max' -%}\n"
+        "<|system|>Reasoning Effort: {{ effective_reasoning_effort | capitalize }}\n"
+        "{% for tc in m.tool_calls %}\n"
+        f"{{{{- '<tool_call>' {concat} tc.name -}}}}\n"
+        "{% set _args = tc.arguments %}"
+        "{% for k, v in _args.items() %}"
+        "<arg_key>{{ k }}</arg_key><arg_value>{{ v }}</arg_value>"
+        "{% endfor %}</tool_call>\n"
+        "{% endfor %}\n"
+        "<|assistant|>{{- '<think>' -}}"
+    )
+
+
 class TestTemplateManagerReasoningDetection(unittest.TestCase):
     def _detect(self, template, vocab):
         force, config = detect_reasoning_pattern(template)
@@ -79,6 +99,23 @@ class TestTemplateManagerReasoningDetection(unittest.TestCase):
         )
         self.assertEqual(parser, "glm45")
 
+    def test_glm53_effort_template_resolves_glm_parsers(self):
+        # Without an enable_thinking toggle the GLM-4.5 rule misses, and the
+        # template used to fall through to deepseek-r1 + the xml_kv fallback
+        # (glm45 tool parser), which cannot read the compact tool-call format.
+        vocab = ["<tool_call>", "<arg_key>", "<arg_value>", "<|user|>", "<|endoftext|>"]
+        for concat in ("+", "~"):
+            template = _glm53_template(concat)
+            with self.subTest(concat=concat):
+                force, config, parser = self._detect(template, vocab)
+                self.assertEqual(parser, "glm45")
+                self.assertEqual(
+                    detect_tool_call_parser(
+                        template, _DummyTokenizer(vocab), config, force
+                    ),
+                    "glm47",
+                )
+
     def test_interns1_detects_enable_thinking_default_true(self):
         template = """
         {% set default_thinking_sys %}...<think>...</think>{% endset %}
@@ -716,6 +753,22 @@ class TestToolCallParserDetection(unittest.TestCase):
                 ["<tool_call>", "<arg_key>", "<arg_value>", "<|endoftext|>"],
                 "glm47",
             ),
+            (
+                "glm47_tilde_concat_tool_call",
+                (
+                    "[gMASK]<sop>\n"
+                    "{% set enable_thinking = enable_thinking if enable_thinking is defined else true %}\n"
+                    "{% for tc in m.tool_calls %}\n"
+                    "{{- '<tool_call>' ~ tc.name -}}\n"
+                    "{% set _args = tc.arguments %}"
+                    "{% for k, v in _args.items() %}"
+                    "<arg_key>{{ k }}</arg_key><arg_value>{{ v }}</arg_value>"
+                    "{% endfor %}</tool_call>\n"
+                    "{% endfor %}"
+                ),
+                ["<tool_call>", "<arg_key>", "<arg_value>", "<|endoftext|>"],
+                "glm47",
+            ),
             (
                 "glm45_newline_tool_call",
                 (
@@ -760,6 +813,9 @@ class TestToolCallParserDetection(unittest.TestCase):
             rule.name: i for i, rule in enumerate(REASONING_PARSER_RULES)
         }
         self.assertLess(reasoning_index["deepseek_v4"], reasoning_index["deepseek_v3"])
+        self.assertLess(
+            reasoning_index["glm45"], reasoning_index["deepseek_r1_think_tags"]
+        )
         self.assertLess(
             reasoning_index["hunyuan"], reasoning_index["deepseek_r1_think_tags"]
         )