feat: use XGrammar V4.1 DSML parameter constraints (#39026)

Co-authored-by: yuchuan <yuchuan.7streams@gmail.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

python/pyproject.toml

@@ -96,7 +96,7 @@ dependencies = [
   "uvicorn",
   "uvloop",
   "watchfiles",
-  "xgrammar==0.2.1",
+  "xgrammar==0.2.7",
   "xxhash",
   "zstandard",
 ]

python/pyproject_cpu.toml

@@ -70,7 +70,7 @@ dependencies = [
   "uvicorn",
   "uvloop",
   "xxhash",
-  "xgrammar==0.2.1",
+  "xgrammar==0.2.7",
   "zstandard",
 ]
 

python/pyproject_npu.toml

@@ -67,7 +67,7 @@ dependencies = [
   "uvicorn",
   "uvloop",
   "xxhash",
-  "xgrammar==0.2.1",
+  "xgrammar==0.2.7",
 ]
 
 [project.optional-dependencies]

python/pyproject_other.toml

@@ -73,7 +73,7 @@ runtime_common = [
   "compressed-tensors",
   "outlines==0.1.11",
   "timm==1.0.16",
-  "xgrammar==0.2.1",
+  "xgrammar==0.2.7",
 ]
 
 # srt_empty: device-agnostic install — pure Python packages only, no torch dependency chain.

python/pyproject_xpu.toml

@@ -69,7 +69,7 @@ dependencies = [
   "uvicorn",
   "xxhash",
   "uvloop",
-  # "xgrammar==0.2.1", xgrammar depends on CUDA PyTorch and Triton only
+  # "xgrammar==0.2.7", xgrammar depends on CUDA PyTorch and Triton only
 ]
 
 [project.optional-dependencies]

python/sglang/srt/arg_groups/field_order.py

@@ -205,6 +205,7 @@ POSITIONAL_FIELD_ORDER = (
     "stat_loggers",
     "constrained_json_whitespace_pattern",
     "constrained_json_disable_any_whitespace",
+    "constrained_json_max_whitespace_cnt",
     "attention_backend",
     "decode_attention_backend",
     "enable_lean_attention",

python/sglang/srt/arg_groups/fields/serving.py

@@ -281,6 +281,10 @@ class Serving(msgspec.Struct):
         bool,
         "(xgrammar and llguidance backends only) Enforce compact representation in JSON constrained output.",
     ] = False
+    constrained_json_max_whitespace_cnt: A[
+        Optional[int],
+        "(xgrammar backend only) Max consecutive whitespace chars allowed in JSON constrained output. None means unbounded.",
+    ] = None
 
     # -------------------------------------------------------------------------
     # Dynamic batch tokenizer

python/sglang/srt/constrained/base_grammar_backend.py

@@ -385,6 +385,7 @@ def create_grammar_backend(
                 vocab_size=vocab_size,
                 model_eos_token_ids=eos_list,
                 any_whitespace=not get_serving().constrained_json_disable_any_whitespace,
+                max_whitespace_cnt=get_serving().constrained_json_max_whitespace_cnt,
             )
         except TokenizerNotSupportedError as e:
             if get_serving().enable_strict_thinking:

python/sglang/srt/constrained/xgrammar_backend.py

@@ -215,6 +215,7 @@ class XGrammarGrammarBackend(BaseGrammarBackend):
         vocab_size: int,
         model_eos_token_ids: Optional[List[int]] = None,
         any_whitespace: bool = True,
+        max_whitespace_cnt: Optional[int] = None,
     ):
         super().__init__()
 
@@ -244,6 +245,7 @@ class XGrammarGrammarBackend(BaseGrammarBackend):
         self.vocab_size = vocab_size
         self.override_stop_tokens = override_stop_tokens
         self.any_whitespace = any_whitespace
+        self.max_whitespace_cnt = max_whitespace_cnt
 
     @property
     def is_support_token_filter(self):
@@ -348,7 +350,9 @@ class XGrammarGrammarBackend(BaseGrammarBackend):
                 schema = json.loads(key_string)
                 validate_xgrammar_json_schema(schema)
                 ctx = self.grammar_compiler.compile_json_schema(
-                    schema=key_string, any_whitespace=self.any_whitespace
+                    schema=key_string,
+                    any_whitespace=self.any_whitespace,
+                    max_whitespace_cnt=self.max_whitespace_cnt,
                 )
 
         except (

python/sglang/srt/function_call/base_format_detector.py

@@ -404,10 +404,8 @@ class BaseFormatDetector(ABC):
                 (the typical case when --reasoning-parser is configured) so
                 only one layer constrains the reasoning section.
             parallel_tool_calls: Whether multiple tool calls may appear in one
-                assistant response. xgrammar's get_model_structural_tag does
-                not expose this knob, so this base implementation ignores it;
-                only detectors that build their own tags (e.g. Kimi K3)
-                honor it.
+                assistant response. Forwarded to XGrammar to constrain the
+                number of tool calls in the generated structural tag.
 
         Returns:
             StructuralTag if this detector supports model-native tags, otherwise None
@@ -427,6 +425,7 @@ class BaseFormatDetector(ABC):
             tools=converted_tools,
             tool_choice=converted_tool_choice,
             reasoning=thinking_mode,
+            parallel_tool_calls=parallel_tool_calls,
         )
 
     def get_auto_tool_call_structural_tag(

python/sglang/srt/function_call/deepseekv32_detector.py

@@ -75,6 +75,7 @@ class DeepSeekV32Detector(BaseFormatDetector):
     tool_calls_block_name = "function_calls"
     invoke_tag_name = "invoke"
     parameter_tag_name = "parameter"
+    strip_string_param_value: bool = True
 
     def __init__(self):
         super().__init__()
@@ -170,7 +171,11 @@ class DeepSeekV32Detector(BaseFormatDetector):
 
             # Convert value based on type
             if param_type == "true":  # string type
-                parameters[param_name] = param_value.strip()
+                parameters[param_name] = (
+                    param_value.strip()
+                    if self.strip_string_param_value
+                    else param_value
+                )
             else:
                 # Try to parse as JSON for other types
                 try:
@@ -195,7 +200,11 @@ class DeepSeekV32Detector(BaseFormatDetector):
             if partial_match and (param_value := partial_match.group(3)):
                 param_name = partial_match.group(1)
                 if partial_match.group(2) == "true":
-                    parameters[param_name] = param_value.strip()
+                    parameters[param_name] = (
+                        param_value.strip()
+                        if self.strip_string_param_value
+                        else param_value
+                    )
                 else:
                     try:
                         parameters[param_name] = _partial_json_loads(

python/sglang/srt/function_call/deepseekv41_detector.py

@@ -1,18 +1,5 @@
-from typing import List, Literal, Optional, Union
+from typing import Optional
 
-from xgrammar.structural_tag import (
-    AnyTextFormat,
-    ConstStringFormat,
-    JSONSchemaFormat,
-    OrFormat,
-    SequenceFormat,
-    TagFormat,
-    TagsWithSeparatorFormat,
-    TriggeredTagsFormat,
-)
-
-from sglang.srt.entrypoints.openai.protocol import Tool, ToolChoice
-from sglang.srt.function_call.base_format_detector import StructuralTag
 from sglang.srt.function_call.deepseekv32_detector import DeepSeekV32Detector
 
 
@@ -25,6 +12,7 @@ class DeepSeekV41Detector(DeepSeekV32Detector):
     tool_calls_block_name = " calls"
     invoke_tag_name = " invoke"
     parameter_tag_name = " parameter"
+    strip_string_param_value: bool = False
 
     # The encoder joins an assistant turn's content and its calls block with a
     # blank line, and renders it even when there is no content.
@@ -32,72 +20,4 @@ class DeepSeekV41Detector(DeepSeekV32Detector):
     think_end_token = "</think>"
 
     def get_structural_tag_name(self) -> Optional[str]:
-        # xgrammar's builtin "deepseek_v4" tag hardcodes the unspaced names,
-        # so the V4.1 tag is assembled in get_structural_tag instead.
-        return None
-
-    def get_structural_tag(
-        self,
-        tools: Union[List[Tool], None] = None,
-        tool_choice: Union[ToolChoice, Literal["auto", "required"]] = "auto",
-        thinking_mode: bool = False,
-        parallel_tool_calls: bool = True,
-    ) -> Optional[StructuralTag]:
-        """The builtin "deepseek_v4" shape with the spaced tag names.
-
-        Bodies are JSON: xgrammar's "deepseek_xml" body style also hardcodes
-        the unspaced "parameter" name, and the V3.2-lineage parser accepts a
-        JSON body inside an invoke.
-        """
-        tools = list(tools or [])
-        if isinstance(tool_choice, ToolChoice):
-            tools = [
-                tool
-                for tool in tools
-                if tool.function.name == tool_choice.function.name
-            ]
-            if len(tools) != 1:
-                return None
-        if not tools:
-            return None
-
-        def invoke_tag(tool: Tool) -> TagFormat:
-            function = tool.function
-            schema = function.parameters if function.strict else True
-            if schema is None:
-                schema = True
-            return TagFormat(
-                begin=f'{self.invoke_start_token} name="{function.name}">',
-                content=JSONSchemaFormat(json_schema=schema),
-                end=f"{self.invoke_end_token}\n",
-            )
-
-        tags = [invoke_tag(tool) for tool in tools]
-        if isinstance(tool_choice, ToolChoice):
-            calls = tags[0]
-        elif parallel_tool_calls:
-            calls = TagsWithSeparatorFormat(tags=tags, separator="", at_least_one=True)
-        else:
-            calls = OrFormat(elements=tags)
-        block_begin = f"{self.bot_token}\n"
-
-        if tool_choice == "auto":
-            body = TriggeredTagsFormat(
-                triggers=[self.bot_token],
-                tags=[TagFormat(begin=block_begin, content=calls, end=self.eot_token)],
-                excludes=["<think>", self.think_end_token],
-            )
-        else:
-            body = SequenceFormat(
-                elements=[
-                    ConstStringFormat(value=self.tool_calls_prefix + block_begin),
-                    calls,
-                    ConstStringFormat(value=self.eot_token),
-                ]
-            )
-        if not thinking_mode:
-            return StructuralTag(format=body)
-        reasoning = TagFormat(
-            begin="", content=AnyTextFormat(), end=self.think_end_token
-        )
-        return StructuralTag(format=SequenceFormat(elements=[reasoning, body]))
+        return "deepseek_v4_1"

python/sglang/srt/function_call/kimik2_detector.py

@@ -468,6 +468,7 @@ class KimiK2Detector(BaseFormatDetector):
             tools=converted_tools,
             tool_choice=converted_tool_choice,
             reasoning=thinking_mode,
+            parallel_tool_calls=parallel_tool_calls,
         )
 
     def get_structural_tag_name(self) -> str: