fix: normalize tool message content for GLM5.1 chat template (#22595)

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -60,6 +60,28 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
+def normalize_tool_content(role: str, content):
+    """Normalize tool message content from OpenAI array format to plain string.
+
+    OpenAI clients may send tool content as a list of content parts
+    (e.g. [{"type":"text","text":"..."}]) but most chat templates expect
+    a plain string for tool messages. Only flatten when ALL items are
+    pure OpenAI text parts; preserve lists containing non-text-type items
+    that some templates intentionally iterate over.
+    """
+    if role != "tool" or not isinstance(content, list):
+        return content
+    parts = content
+    is_openai_text_parts = all(
+        (isinstance(p, dict) and p.get("type") == "text") or isinstance(p, str)
+        for p in parts
+    )
+    if is_openai_text_parts:
+        text_parts = [p.get("text", "") if isinstance(p, dict) else p for p in parts]
+        return " ".join(text_parts)
+    return content
+
+
 def _extract_max_dynamic_patch(request: ChatCompletionRequest):
     img_vals = []
     vid_vals = []
@@ -457,6 +479,10 @@ class OpenAIServingChat(OpenAIServingBase):
                     modalities,
                 )
 
+                processed_msg["content"] = normalize_tool_content(
+                    processed_msg["role"], processed_msg.get("content")
+                )
+
                 # per the Transformers docs & maintainers, tool call arguments in
                 # assistant-role messages with tool_calls need to be dicts not JSON str -
                 # this is how tool-use chat templates will expect them moving forwards

test/registered/openai_server/basic/test_serving_chat.py

@@ -19,7 +19,10 @@ from sglang.srt.entrypoints.openai.protocol import (
     ChatCompletionRequest,
     MessageProcessingResult,
 )
-from sglang.srt.entrypoints.openai.serving_chat import OpenAIServingChat
+from sglang.srt.entrypoints.openai.serving_chat import (
+    OpenAIServingChat,
+    normalize_tool_content,
+)
 from sglang.srt.managers.io_struct import GenerateReqInput
 from sglang.srt.utils import get_or_create_event_loop
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
@@ -982,5 +985,42 @@ class TestProcessToolCallsWithRequiredToolChoice(unittest.TestCase):
         self.assertIsNone(tool_calls)
 
 
+class TestNormalizeToolContent(unittest.TestCase):
+    """Unit tests for normalize_tool_content()."""
+
+    def test_openai_text_parts_flattened(self):
+        result = normalize_tool_content("tool", [{"type": "text", "text": "10525"}])
+        self.assertEqual(result, "10525")
+
+    def test_multiple_text_parts_joined(self):
+        result = normalize_tool_content(
+            "tool",
+            [{"type": "text", "text": "hello"}, {"type": "text", "text": "world"}],
+        )
+        self.assertEqual(result, "hello world")
+
+    def test_non_text_part_list_preserved(self):
+        content = [{"name": "func", "output": "result"}]
+        result = normalize_tool_content("tool", content)
+        self.assertIs(result, content)
+
+    def test_string_content_unchanged(self):
+        self.assertEqual(normalize_tool_content("tool", "hello"), "hello")
+
+    def test_empty_list_returns_empty_string(self):
+        self.assertEqual(normalize_tool_content("tool", []), "")
+
+    def test_non_tool_role_unchanged(self):
+        content = [{"type": "text", "text": "hi"}]
+        result = normalize_tool_content("user", content)
+        self.assertIs(result, content)
+
+    def test_mixed_str_and_dict_parts(self):
+        result = normalize_tool_content(
+            "tool", ["plain", {"type": "text", "text": "rich"}]
+        )
+        self.assertEqual(result, "plain rich")
+
+
 if __name__ == "__main__":
     unittest.main(verbosity=2)