Adding tool calling and reasoning parser support for Intern-S1 (#14866)

Co-authored-by: KennyYao2001 <kennyy@andrew.cmu.edu>
Co-authored-by: Cursor AI <cursor@example.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Changyi Yang <112288487+ChangyiYang@users.noreply.github.com>

python/sglang/srt/constrained/base_grammar_backend.py

@@ -226,17 +226,29 @@ def create_grammar_backend(
             whitespace_pattern=server_args.constrained_json_whitespace_pattern,
         )
     elif name == "xgrammar":
-        from sglang.srt.constrained.xgrammar_backend import XGrammarGrammarBackend
+        from sglang.srt.constrained.xgrammar_backend import (
+            TokenizerNotSupportedError,
+            XGrammarGrammarBackend,
+        )
 
         # Convert Set[int] to List[int] if needed
         eos_list = list(eos_token_ids) if eos_token_ids else None
 
-        grammar_backend = XGrammarGrammarBackend(
-            tokenizer,
-            vocab_size=vocab_size,
-            model_eos_token_ids=eos_list,
-            any_whitespace=not server_args.constrained_json_disable_any_whitespace,
-        )
+        try:
+            grammar_backend = XGrammarGrammarBackend(
+                tokenizer,
+                vocab_size=vocab_size,
+                model_eos_token_ids=eos_list,
+                any_whitespace=not server_args.constrained_json_disable_any_whitespace,
+            )
+        except TokenizerNotSupportedError as e:
+            logger.warning(
+                f"Grammar backend disabled because tokenizer is not supported by XGrammar: {e}. "
+                "Falling back to grammar_backend='none'. "
+                "Structured outputs (JSON schema, regex, EBNF) will not be available."
+            )
+            server_args.grammar_backend = "none"
+            return None
     elif name == "llguidance":
         from sglang.srt.constrained.llguidance_backend import GuidanceBackend
 

python/sglang/srt/constrained/xgrammar_backend.py

@@ -166,6 +166,12 @@ class XGrammarGrammar(BaseGrammarObject):
         return f"XGrammarGrammar({self.key_string=}, {self.accepted_tokens=}, {self.current_token=})"
 
 
+class TokenizerNotSupportedError(Exception):
+    """Raised when tokenizer is not supported by XGrammar backend."""
+
+    pass
+
+
 class XGrammarGrammarBackend(BaseGrammarBackend):
     def __init__(
         self,
@@ -182,14 +188,21 @@ class XGrammarGrammarBackend(BaseGrammarBackend):
 
             if tokenizer_info is None:
                 # Not supported tokenizer
-                return
+                raise TokenizerNotSupportedError(
+                    f"Tokenizer type {type(tokenizer).__name__} is not supported by XGrammar"
+                )
         else:
             # Create TokenizerInfo with model's EOS tokens as the authoritative stop tokens
             # This ensures consistency between what the model considers EOS and what XGrammar uses
-            tokenizer_info = TokenizerInfo.from_huggingface(
-                tokenizer, vocab_size=vocab_size, stop_token_ids=model_eos_token_ids
-            )
-            override_stop_tokens = None
+            try:
+                tokenizer_info = TokenizerInfo.from_huggingface(
+                    tokenizer, vocab_size=vocab_size, stop_token_ids=model_eos_token_ids
+                )
+                override_stop_tokens = None
+            except Exception as e:
+                raise TokenizerNotSupportedError(
+                    f"Failed to create XGrammar TokenizerInfo from tokenizer: {e}"
+                )
 
         self.grammar_compiler = GrammarCompiler(tokenizer_info=tokenizer_info)
         self.vocab_size = vocab_size

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

@@ -1069,8 +1069,8 @@ class OpenAIServingChat(OpenAIServingBase):
                 request.chat_template_kwargs is not None
                 and request.chat_template_kwargs.get("thinking") is True
             )
-        if self.reasoning_parser in ["qwen3", "glm45", "nano_v3"]:
-            # qwen3, glm45, and nano_v3 are reasoning by default
+        if self.reasoning_parser in ["qwen3", "glm45", "nano_v3", "interns1"]:
+            # qwen3, glm45, nano_v3, and interns1 are reasoning by default
             return (
                 not request.chat_template_kwargs
                 or request.chat_template_kwargs.get("enable_thinking", True) is True

python/sglang/srt/function_call/function_call_parser.py

@@ -16,6 +16,7 @@ from sglang.srt.function_call.deepseekv31_detector import DeepSeekV31Detector
 from sglang.srt.function_call.deepseekv32_detector import DeepSeekV32Detector
 from sglang.srt.function_call.glm4_moe_detector import Glm4MoeDetector
 from sglang.srt.function_call.gpt_oss_detector import GptOssDetector
+from sglang.srt.function_call.internlm_detector import InternlmDetector
 from sglang.srt.function_call.kimik2_detector import KimiK2Detector
 from sglang.srt.function_call.llama32_detector import Llama32Detector
 from sglang.srt.function_call.minimax_m2 import MinimaxM2Detector
@@ -54,6 +55,7 @@ class FunctionCallParser:
         "qwen3_coder": Qwen3CoderDetector,
         "step3": Step3Detector,
         "minimax-m2": MinimaxM2Detector,
+        "interns1": InternlmDetector,
     }
 
     def __init__(self, tools: List[Tool], tool_call_parser: str):

python/sglang/srt/function_call/internlm_detector.py

@@ -0,0 +1,248 @@
+# modified from https://github.com/InternLM/lmdeploy/blob/main/lmdeploy/serve/openai/tool_parser/internlm2_parser.py
+
+import json
+import logging
+import re
+from typing import List
+
+from sglang.srt.entrypoints.openai.protocol import Tool
+from sglang.srt.environ import envs
+from sglang.srt.function_call.base_format_detector import BaseFormatDetector
+from sglang.srt.function_call.core_types import (
+    StreamingParseResult,
+    StructureInfo,
+    ToolCallItem,
+    _GetInfoFunc,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class InternlmDetector(BaseFormatDetector):
+    """
+    Detector for InternLM2/Intern-S1 model function call format.
+
+    The InternLM format uses special tokens to delimit function calls
+    with JSON for arguments.
+
+    Format Structure:
+    ```
+    text<|action_start|> <|plugin|>
+    {json}<|action_end|>
+    ```
+
+    Examples:
+    ```
+    What's the weather like?<|action_start|> <|plugin|>
+    {"name": "get_weather", "parameters": {"location": "Tokyo"}}<|action_end|>
+    ```
+
+    Key Components:
+    - Tool Call Start: `<|action_start|> <|plugin|>`
+    - Tool Call End: `<|action_end|>`
+    - Arguments: JSON object with `name` and `parameters`/`arguments`
+    - Supports multiple sequential tool calls in both streaming and non-streaming modes
+
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.bot_token = "<|action_start|> <|plugin|>"
+        self.eot_token = "<|action_end|>"
+        self.position = 0
+
+    def has_tool_call(self, text: str) -> bool:
+        """Check if the text contains an InternLM format tool call."""
+        has_call = self.bot_token in text
+        return has_call
+
+    def get_arguments(self, obj):
+        """Extract arguments from object, supporting both 'parameters' and 'arguments' keys."""
+        if "parameters" in obj:
+            return obj.get("parameters")
+        elif "arguments" in obj:
+            return obj.get("arguments")
+        return None
+
+    def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
+        """
+        One-time parsing: Detects and parses tool calls in the provided text.
+        Supports multiple tool calls in the format:
+        <|action_start|> <|plugin|>\n{JSON}<|action_end|>
+
+        :param text: The complete text to parse.
+        :param tools: List of available tools.
+        :return: StreamingParseResult with normal text and parsed tool calls.
+        """
+
+        # Find the first occurrence of tool call marker to extract normal text
+        idx = text.find(self.bot_token)
+        normal_text = text[:idx].strip() if idx != -1 else text
+
+        if self.bot_token not in text:
+            logger.warning("[InternLM Tool Call] No tool call markers found in text")
+            return StreamingParseResult(normal_text=normal_text, calls=[])
+
+        # Use regex to find all tool call blocks
+        # Pattern matches: {self.bot_token}{...}{self.eot_token}
+        tool_call_pattern = (
+            rf"{re.escape(self.bot_token)}\s*(.*?){re.escape(self.eot_token)}"
+        )
+        matches = re.findall(tool_call_pattern, text, re.DOTALL)
+
+        if not matches:
+            logger.warning("[InternLM Tool Call] No complete tool call blocks found")
+            return StreamingParseResult(normal_text=text, calls=[])
+
+        logger.info(f"[InternLM Tool Call] Found {len(matches)} tool call(s)")
+
+        calls = []
+        tool_indices = self._get_tool_indices(tools)
+
+        try:
+            for idx, action_json in enumerate(matches):
+                action_json = action_json.strip()
+
+                try:
+                    # Parse the JSON
+                    action_dict = json.loads(action_json)
+                    name = action_dict.get("name")
+                    parameters = self.get_arguments(action_dict)
+
+                    if not parameters:
+                        parameters = {}
+
+                    logger.info(
+                        f"[InternLM Tool Call] Parsed tool call #{idx+1}: name={name}, "
+                        f"parameters={json.dumps(parameters, ensure_ascii=False)}"
+                    )
+
+                    # Validate tool name
+                    if not (name and name in tool_indices):
+                        logger.warning(
+                            f"[InternLM Tool Call] Model attempted to call undefined function: {name}, "
+                            f"available_tools={list(tool_indices.keys())}"
+                        )
+                        if not envs.SGLANG_FORWARD_UNKNOWN_TOOLS.get():
+                            continue  # Skip this tool call
+
+                    # Create tool call item and add to list
+                    tool_call = ToolCallItem(
+                        tool_index=tool_indices[name],
+                        name=name,
+                        parameters=json.dumps(parameters, ensure_ascii=False),
+                    )
+                    calls.append(tool_call)
+
+                except json.JSONDecodeError as e:
+                    logger.error(
+                        f"[InternLM Tool Call] Failed to parse JSON for tool call #{idx+1}: {e}"
+                    )
+                    continue
+
+            logger.info(
+                f"[InternLM Tool Call] Successfully parsed {len(calls)} tool call(s), "
+                f"normal_text_length={len(normal_text)}"
+            )
+            return StreamingParseResult(normal_text=normal_text, calls=calls)
+
+        except Exception as e:
+            logger.error(
+                f"[InternLM Tool Call] Error in detect_and_parse: {e}", exc_info=True
+            )
+            return StreamingParseResult(normal_text=text, calls=[])
+
+    def parse_streaming_increment(
+        self, new_text: str, tools: List[Tool]
+    ) -> StreamingParseResult:
+        """
+        Streaming incremental parsing for InternLM format.
+
+        Supports a single tool call in streaming mode.
+        """
+        self._buffer += new_text
+        current_text = self._buffer
+
+        # Check if we don't have a tool call start marker
+        start = current_text.find(self.bot_token)
+        if start == -1:
+            # No tool call marker found
+            # If we've already processed tool calls, don't return text again
+            if self.current_tool_id > 0:
+                self._buffer = ""
+                return StreamingParseResult(normal_text="")
+
+            # Check if buffer could be partial start of bot_token
+            if not self._ends_with_partial_token(current_text, self.bot_token):
+                # Not a partial match, return as normal text
+                normal_text = current_text
+                self._buffer = ""
+                # Clean up any stray end tokens
+                if self.eot_token in normal_text:
+                    normal_text = normal_text.replace(self.eot_token, "")
+                return StreamingParseResult(normal_text=normal_text)
+            else:
+                # Might be partial start token, keep buffering
+                return StreamingParseResult()
+
+        # Check if we have a complete tool call (with end marker)
+        end = current_text.find(self.eot_token)
+        if end != -1:
+            # We have a complete tool call
+            # Initialize state if this is the first tool call
+            if self.current_tool_id == -1:
+                self.current_tool_id = 0
+                self.prev_tool_call_arr = []
+                self.streamed_args_for_tool = [""]
+
+            # Ensure we have enough entries in our tracking arrays
+            while len(self.prev_tool_call_arr) <= self.current_tool_id:
+                self.prev_tool_call_arr.append({})
+            while len(self.streamed_args_for_tool) <= self.current_tool_id:
+                self.streamed_args_for_tool.append("")
+
+            # Use detect_and_parse on the complete tool call
+            complete_section = current_text[: end + len(self.eot_token)]
+            result = self.detect_and_parse(complete_section, tools=tools)
+
+            if result.calls:
+                # Update the tool call index
+                result.calls[0].tool_index = self.current_tool_id
+                # Store the parsed tool call for reference
+                self.prev_tool_call_arr[self.current_tool_id] = {
+                    "name": result.calls[0].name,
+                    "arguments": json.loads(result.calls[0].parameters),
+                }
+                self.streamed_args_for_tool[self.current_tool_id] = result.calls[
+                    0
+                ].parameters
+                # Increment tool ID for next tool call
+                self.current_tool_id += 1
+
+            # Remove the completed tool call from buffer
+            self._buffer = current_text[end + len(self.eot_token) :]
+            return result
+
+        # We have bot_token but no eot_token yet - handle partial tool call streaming
+        # Extract normal text before the tool call
+        normal_text = current_text[:start]
+        # Keep the tool call part in buffer
+        self._buffer = current_text[start:]
+        return StreamingParseResult(normal_text=normal_text)
+
+    def structure_info(self) -> _GetInfoFunc:
+        """
+        Return structure information for constrained generation.
+
+        For InternLM format, the structure is:
+        - begin: <|action_start|> <|plugin|>\n
+        - end: <|action_end|>
+        - trigger: the begin token
+        """
+        return lambda name: StructureInfo(
+            begin='<|action_start|> <|plugin|>\n{"name": "'
+            + name
+            + '", "parameters": ',
+            end="}<|action_end|>",
+            trigger="<|action_start|> <|plugin|>",
+        )

python/sglang/srt/parser/reasoning_parser.py

@@ -314,6 +314,7 @@ class ReasoningParser:
         "minimax-append-think": MiniMaxAppendThinkDetector,
         "step3": DeepSeekR1Detector,
         "nano_v3": NanoV3Detector,
+        "interns1": Qwen3Detector,
     }
 
     def __init__(