[Tool Call][DSV32] Streamline function call parameters (#14750)

Signed-off-by: Muqi Li <muqi1029@gmail.com>

python/sglang/srt/function_call/deepseekv32_detector.py

@@ -2,6 +2,8 @@ import json
 import logging
 import re
 
+from partial_json_parser.core.options import Allow
+
 from sglang.srt.entrypoints.openai.protocol import Tool
 from sglang.srt.function_call.base_format_detector import BaseFormatDetector
 from sglang.srt.function_call.core_types import (
@@ -10,7 +12,7 @@ from sglang.srt.function_call.core_types import (
     ToolCallItem,
     _GetInfoFunc,
 )
-from sglang.srt.function_call.utils import _find_common_prefix
+from sglang.srt.function_call.utils import _find_common_prefix, _partial_json_loads
 
 logger = logging.getLogger(__name__)
 
@@ -82,11 +84,12 @@ class DeepSeekV32Detector(BaseFormatDetector):
         self.invoke_regex = (
             r'<｜DSML｜invoke\s+name="([^"]+)"\s*>(.*?)(</｜DSML｜invoke>|$)'
         )
+        self.prefix_parameter_end_call = ["</", "｜DSML｜", "parameter"]
         self.current_tool_id = -1
 
     def has_tool_call(self, text: str) -> bool:
         """Check if the text contains a deepseek v32 format tool call."""
-        return self.bot_token in text
+        return self.bot_token in text or "<｜DSML｜invoke" in text
 
     def _parse_parameters_from_xml(
         self, invoke_content: str, allow_partial: bool = False
@@ -139,16 +142,25 @@ class DeepSeekV32Detector(BaseFormatDetector):
         # If allowed, try to parse a partial parameter at the end
         if allow_partial:
             remaining_content = invoke_content[last_match_end:]
+
+            # Remove incomplete parameter_end_call prefix in case they are captured by param
+            for token in reversed(self.prefix_parameter_end_call):
+                remaining_content = remaining_content.rstrip(token)
+
             # Match start of a parameter tag + value (potentially incomplete)
             # Regex: <tag name="..." string="...">VALUE... (no end tag)
             partial_match = re.search(
                 self.partial_parameter_regex, remaining_content, re.DOTALL
             )
 
-            if partial_match:
+            if partial_match and (param_value := partial_match.group(3)):
                 param_name = partial_match.group(1)
-                param_value = partial_match.group(3)
+                if partial_match.group(2) == "true":
-                parameters[param_name] = param_value
+                    parameters[param_name] = param_value.strip()
+                else:
+                    parameters[param_name] = _partial_json_loads(
+                        param_value, Allow.ALL
+                    )[0]
 
         return parameters
 
@@ -206,13 +218,6 @@ class DeepSeekV32Detector(BaseFormatDetector):
         self._buffer += new_text
         current_text = self._buffer
 
-        # Check if we have a tool call or any DSML-related content
-        # Key insight: DSML tags contain distinctive markers like "｜DSML｜"
-        # If we see these markers anywhere, we should keep buffering
-        has_tool_call = (
-            self.bot_token in current_text or "<｜DSML｜invoke" in current_text
-        )
-
         # Check if buffer contains any DSML markers or ends with potential tag prefix
         # This handles partial/streaming DSML content
         dsml_markers = ["｜DSML｜", "<｜", "</｜"]
@@ -224,7 +229,11 @@ class DeepSeekV32Detector(BaseFormatDetector):
             current_text.rstrip().endswith(prefix) for prefix in dsml_prefixes
         )
 
-        if not has_tool_call and not potentially_dsml and not ends_with_prefix:
+        if (
+            not self.has_tool_call(current_text)
+            and not potentially_dsml
+            and not ends_with_prefix
+        ):
             self._buffer = ""
             for e_token in [self.eot_token, self.invoke_end_token]:
                 if e_token in current_text:
@@ -241,7 +250,6 @@ class DeepSeekV32Detector(BaseFormatDetector):
                     string=current_text,
                     flags=re.DOTALL,
                 )
-
                 if not invoke_match:
                     break
 
@@ -342,5 +350,5 @@ class DeepSeekV32Detector(BaseFormatDetector):
         return lambda name: StructureInfo(
             begin=f'<｜DSML｜invoke name="{name}">',
             end="</｜DSML｜invoke>",
-            trigger=f'<｜DSML｜invoke name="{name}">',
+            trigger=f"<｜DSML｜invoke",
         )

test/registered/function_call/test_function_call_parser.py

@@ -1159,6 +1159,10 @@ class TestDeepSeekV32Detector(unittest.TestCase):
             ),
         ]
         self.detector = DeepSeekV32Detector()
+        from transformers import AutoTokenizer
+
+        self.tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3.2")
+        self.interval = 1
 
     def test_detect_and_parse_xml_format(self):
         """Test parsing standard XML format (DSML)"""
@@ -1236,12 +1240,19 @@ class TestDeepSeekV32Detector(unittest.TestCase):
         text = """<｜DSML｜function_calls>
             <｜DSML｜invoke name="get_favorite_tourist_spot">
                 <｜DSML｜parameter name="city" string="true">San Francisco</｜DSML｜parameter>
+                <｜DSML｜parameter name="another_city" string="true">London</｜DSML｜parameter>
+                <｜DSML｜parameter name="topn" string="false">10</｜DSML｜parameter>
+                <｜DSML｜parameter name="obj" string="false">{"name": "John", "age": 30}</｜DSML｜parameter>
             </｜DSML｜invoke>
         </｜DSML｜function_calls>"""
 
-        chunks = [text[i : i + 5] for i in range(0, len(text), 5)]
+        input_ids = self.tokenizer.encode(text, add_special_tokens=False)
+        chunk_ids = [
+            input_ids[i : i + self.interval]
+            for i in range(0, len(input_ids), self.interval)
+        ]
+        chunks = [self.tokenizer.decode(chunk_id) for chunk_id in chunk_ids]
 
-        accumulated_calls = []
         tool_calls_by_index = {}
 
         for chunk in chunks:
@@ -1263,15 +1274,12 @@ class TestDeepSeekV32Detector(unittest.TestCase):
 
         self.assertEqual(len(tool_calls_by_index), 1)
         self.assertEqual(tool_calls_by_index[0]["name"], "get_favorite_tourist_spot")
-        # Note: The detector might accumulate partial JSON string which is valid,
+        params = json.loads(tool_calls_by_index[0]["parameters"])
-        # but for XML format it constructs JSON at the end.
+        self.assertEqual(params["city"], "San Francisco")
-        # Let's check if the final parameters parse correctly.
+        self.assertEqual(params["another_city"], "London")
-        try:
+        self.assertEqual(params["topn"], 10)
-            params = json.loads(tool_calls_by_index[0]["parameters"])
+        self.assertEqual(params["obj"]["name"], "John")
-            self.assertEqual(params["city"], "San Francisco")
+        self.assertEqual(params["obj"]["age"], 30)
-        except json.JSONDecodeError:
-            # In streaming XML, parameters might be constructed differently or incrementally
-            pass
 
     def test_streaming_json_format(self):
         """Test streaming parsing of JSON format"""
@@ -1283,7 +1291,12 @@ class TestDeepSeekV32Detector(unittest.TestCase):
             </｜DSML｜invoke>
         </｜DSML｜function_calls>"""
 
-        chunks = [text[i : i + 5] for i in range(0, len(text), 5)]
+        input_ids = self.tokenizer.encode(text, add_special_tokens=False)
+        chunk_ids = [
+            input_ids[i : i + self.interval]
+            for i in range(0, len(input_ids), self.interval)
+        ]
+        chunks = [self.tokenizer.decode(chunk_id) for chunk_id in chunk_ids]
 
         tool_calls_by_index = {}
 
@@ -1370,7 +1383,12 @@ class TestDeepSeekV32Detector(unittest.TestCase):
         self.detector = DeepSeekV32Detector()
 
         # Simulate streaming by splitting into small chunks
-        chunks = [text[i : i + 5] for i in range(0, len(text), 5)]
+        input_ids = self.tokenizer.encode(text, add_special_tokens=False)
+        chunk_ids = [
+            input_ids[i : i + self.interval]
+            for i in range(0, len(input_ids), self.interval)
+        ]
+        chunks = [self.tokenizer.decode(chunk_id) for chunk_id in chunk_ids]
 
         tool_calls_by_index = {}
 
@@ -1425,7 +1443,12 @@ class TestDeepSeekV32Detector(unittest.TestCase):
         # Reset detector state
         self.detector = DeepSeekV32Detector()
 
-        chunks = [text[i : i + 5] for i in range(0, len(text), 5)]
+        input_ids = self.tokenizer.encode(text, add_special_tokens=False)
+        chunk_ids = [
+            input_ids[i : i + self.interval]
+            for i in range(0, len(input_ids), self.interval)
+        ]
+        chunks = [self.tokenizer.decode(chunk_id) for chunk_id in chunk_ids]
 
         tool_calls_by_index = {}