[GLM-4.7] GLM-4.7 Tool Parser and Doc Update (#15333)

This commit is contained in:
Yuxuan Zhang
2025-12-20 12:30:44 +08:00
committed by GitHub
parent c0f9b51992
commit b82c7a0ae7
7 changed files with 849 additions and 434 deletions

View File

@@ -15,6 +15,7 @@ from sglang.srt.function_call.deepseekv3_detector import DeepSeekV3Detector
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.glm47_moe_detector import Glm47MoeDetector
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
@@ -46,6 +47,7 @@ class FunctionCallParser:
"deepseekv32": DeepSeekV32Detector,
"glm": Glm4MoeDetector,
"glm45": Glm4MoeDetector,
"glm47": Glm47MoeDetector,
"gpt-oss": GptOssDetector,
"kimi_k2": KimiK2Detector,
"llama3": Llama32Detector,

View File

@@ -0,0 +1,584 @@
import ast
import json
import logging
import re
from enum import Enum
from typing import Any, Dict, List, Optional, Tuple
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 (
StreamingParseResult,
ToolCallItem,
_GetInfoFunc,
)
logger = logging.getLogger(__name__)
class StreamState(str, Enum):
"""State machine states for XML to JSON streaming conversion."""
INIT = "INIT"
BETWEEN = "BETWEEN"
IN_KEY = "IN_KEY"
WAITING_VALUE = "WAITING_VALUE"
IN_VALUE = "IN_VALUE"
def get_argument_type(
func_name: str, arg_key: str, defined_tools: List[Tool]
) -> Optional[str]:
"""Get the expected type of a function argument from tool definitions.
Args:
func_name: Name of the function/tool
arg_key: Name of the argument
defined_tools: List of available tools
Returns:
The type string (e.g., 'string', 'number', 'object') or None if not found
"""
name2tool = {tool.function.name: tool for tool in defined_tools}
if func_name not in name2tool:
return None
tool = name2tool[func_name]
properties = (tool.function.parameters or {}).get("properties", {})
if not isinstance(properties, dict):
properties = {}
if arg_key not in properties:
return None
return properties[arg_key].get("type", None)
def _convert_to_number(value: str) -> Any:
"""Convert string to appropriate number type (int or float).
Args:
value: String value to convert
Returns:
Converted number or original string if conversion fails
"""
try:
if "." in value or "e" in value.lower():
return float(value)
else:
return int(value)
except (ValueError, AttributeError):
return value
def parse_arguments(
json_value: str, arg_type: Optional[str] = None
) -> Tuple[Any, bool]:
"""Parse argument value with multiple fallback strategies.
Args:
json_value: Raw string value to parse
arg_type: Expected type hint ('string', 'number', 'object', etc.)
Returns:
Tuple of (parsed_value, is_valid_json)
"""
# Strategy 1: Direct JSON parsing
try:
parsed_value = json.loads(json_value)
# Type coercion for number type
if arg_type == "number" and isinstance(parsed_value, str):
parsed_value = _convert_to_number(parsed_value)
return parsed_value, True
except (json.JSONDecodeError, ValueError):
pass
# Strategy 2: Unescape and parse
try:
wrapped = json.loads('{"tmp": "' + json_value + '"}')
parsed_value = json.loads(wrapped["tmp"])
if arg_type == "number" and isinstance(parsed_value, str):
parsed_value = _convert_to_number(parsed_value)
return parsed_value, True
except (json.JSONDecodeError, ValueError, KeyError):
pass
# Strategy 3: ast.literal_eval
try:
parsed_value = ast.literal_eval(json_value)
return parsed_value, True
except (ValueError, SyntaxError):
pass
# Strategy 4: Treat as string
try:
quoted_value = json.dumps(str(json_value))
return json.loads(quoted_value), True
except (json.JSONDecodeError, ValueError):
return json_value, False
class Glm47MoeDetector(BaseFormatDetector):
"""
Detector for GLM-4.7 and GLM-5 models.
Assumes function call format:
<tool_call>get_weather<arg_key>city</arg_key><arg_value>北京</arg_value><arg_key>date</arg_key><arg_value>2024-06-27</arg_value></tool_call><tool_call>get_weather<arg_key>city</arg_key><arg_value>上海</arg_value><arg_key>date</arg_key><arg_value>2024-06-27</arg_value></tool_call>
"""
def __init__(self):
super().__init__()
self.bot_token = "<tool_call>"
self.eot_token = "</tool_call>"
self.func_call_regex = r"<tool_call>.*?</tool_call>"
self.func_detail_regex = re.compile(
r"<tool_call>(.*?)(<arg_key>.*?)?</tool_call>", re.DOTALL
)
self.func_arg_regex = re.compile(
r"<arg_key>(.*?)</arg_key>(?:\\n|\s)*<arg_value>(.*?)</arg_value>",
re.DOTALL,
)
self._last_arguments = ""
self.current_tool_id = -1
self.current_tool_name_sent = False
self._streamed_raw_length = 0
self._reset_streaming_state()
def _reset_streaming_state(self) -> None:
"""Reset the streaming state machine for a new tool call."""
self._stream_state = StreamState.INIT
self._current_key = ""
self._current_value = ""
self._xml_tag_buffer = ""
self._is_first_param = True
self._value_started = False
self._cached_value_type: Optional[str] = (
None # Cache the value type for consistency
)
def has_tool_call(self, text: str) -> bool:
"""Check if the text contains a glm-4.5 / glm-4.6 format tool call."""
return self.bot_token in text
def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
"""
One-time parsing: Detects and parses tool calls in the provided text.
:param text: The complete text to parse.
:param tools: List of available tools.
:return: ParseResult indicating success or failure, consumed text, leftover text, and parsed calls.
"""
idx = text.find(self.bot_token)
normal_text = text[:idx].strip() if idx != -1 else text
if self.bot_token not in text:
return StreamingParseResult(normal_text=normal_text, calls=[])
match_result_list = re.findall(self.func_call_regex, text, re.DOTALL)
calls = []
try:
for match_result in match_result_list:
# Get function name
func_detail = self.func_detail_regex.search(match_result)
func_name = func_detail.group(1)
func_args = func_detail.group(2)
arguments = {}
if func_args:
pairs = self.func_arg_regex.findall(func_args)
# Parse arguments using shared method
arguments = self._parse_argument_pairs(pairs, func_name, tools)
# construct match_result for parse_base_json
match_result = {"name": func_name, "parameters": arguments}
calls.extend(self.parse_base_json(match_result, tools))
return StreamingParseResult(normal_text=normal_text, calls=calls)
except Exception as e:
logger.error(f"Error in detect_and_parse: {e}", exc_info=True)
# return the normal text if parsing fails
return StreamingParseResult(normal_text=text)
def _get_value_type(self, func_name: str, key: str, tools: List[Tool]) -> str:
"""Get parameter type from tool definition, with fallback to auto-detection.
Args:
func_name: Name of the function
key: Parameter name
tools: List of available tools
Returns:
Type string: 'string', 'number', or 'object'
"""
arg_type = get_argument_type(func_name, key, tools)
if arg_type:
return arg_type
# Auto-detect type from value (best effort)
first_chars = self._current_value.strip()[:10] if self._current_value else ""
if first_chars:
first_char = first_chars[0]
if first_char.isdigit() or first_char in ["-", "."]:
return "number"
elif first_char in ["{", "["]:
return "object"
return "string"
def _format_value_complete(self, value: str, value_type: str) -> str:
"""Format complete value based on type.
Args:
value: Raw value string
value_type: Expected type ('string', 'number', 'object')
Returns:
Properly formatted JSON value string
"""
if value_type == "string":
# Ensure proper JSON string formatting with quotes
return json.dumps(value, ensure_ascii=False)
elif value_type == "number":
try:
num = _convert_to_number(value.strip())
return str(num)
except (ValueError, AttributeError):
# Fallback to string if not a valid number
logger.warning(
f"Failed to parse '{value}' as number, treating as string"
)
return json.dumps(str(value), ensure_ascii=False)
else:
# For object/array types, return as-is (should already be valid JSON)
return value
def _process_xml_to_json_streaming(
self, raw_increment: str, func_name: str, tools: List[Tool]
) -> str:
"""Convert XML increment to JSON streaming output using state machine.
This method processes XML fragments character by character and converts them
to JSON format incrementally. It maintains state across calls to handle
partial XML tags and values.
Args:
raw_increment: New XML content to process
func_name: Name of the function being called
tools: List of available tools for type inference
Returns:
JSON string increment to append to the output
"""
json_output = ""
for char in raw_increment:
self._xml_tag_buffer += char
if self._stream_state in [StreamState.INIT, StreamState.BETWEEN]:
if self._xml_tag_buffer.endswith("<arg_key>"):
self._stream_state = StreamState.IN_KEY
self._current_key = ""
self._xml_tag_buffer = ""
json_output += "{" if self._is_first_param else ", "
self._is_first_param = False
elif self._stream_state == StreamState.IN_KEY:
if self._xml_tag_buffer.endswith("</arg_key>"):
self._current_key = self._xml_tag_buffer[:-10].strip()
self._xml_tag_buffer = ""
self._stream_state = StreamState.WAITING_VALUE
json_output += (
json.dumps(self._current_key, ensure_ascii=False) + ": "
)
elif self._stream_state == StreamState.WAITING_VALUE:
if self._xml_tag_buffer.endswith("<arg_value>"):
self._stream_state = StreamState.IN_VALUE
self._current_value = ""
self._xml_tag_buffer = ""
self._value_started = False
# Determine and cache the value type at the start
self._cached_value_type = self._get_value_type(
func_name, self._current_key, tools
)
elif self._stream_state == StreamState.IN_VALUE:
if self._xml_tag_buffer.endswith("</arg_value>"):
final_value = self._xml_tag_buffer[:-12]
self._current_value += final_value
# Use cached value type for consistency
value_type = self._cached_value_type or "string"
if self._value_started:
# Output any remaining content
if final_value:
if value_type == "string":
json_output += json.dumps(
final_value, ensure_ascii=False
)[1:-1]
else:
json_output += final_value
# Always output closing quote for string type when value was started
if value_type == "string":
json_output += '"'
else:
# Value was never started (empty or complete in one chunk)
json_output += self._format_value_complete(
self._current_value, value_type
)
self._xml_tag_buffer = ""
self._stream_state = StreamState.BETWEEN
self._current_value = ""
self._value_started = False
self._cached_value_type = None # Reset cached type
else:
closing_tag = "</arg_value>"
is_potential_closing = len(self._xml_tag_buffer) <= len(
closing_tag
) and closing_tag.startswith(self._xml_tag_buffer)
if not is_potential_closing:
content = self._xml_tag_buffer
# Use cached value type for consistency
value_type = self._cached_value_type or "string"
if value_type == "string":
if not self._value_started:
json_output += '"'
self._value_started = True
if content:
json_output += json.dumps(content, ensure_ascii=False)[
1:-1
]
self._current_value += content
self._xml_tag_buffer = ""
elif value_type == "number":
if content:
if not self._value_started:
self._value_started = True
json_output += content
self._current_value += content
self._xml_tag_buffer = ""
else:
# For object/array types, output as-is
if content:
if not self._value_started:
self._value_started = True
json_output += content
self._current_value += content
self._xml_tag_buffer = ""
return json_output
def parse_streaming_increment(
self, new_text: str, tools: List[Tool]
) -> StreamingParseResult:
"""
Streaming incremental parsing tool calls for GLM-4.5 and GLM-4.6 format.
Uses a state machine to convert XML to JSON incrementally for true character-by-character streaming.
Outputs JSON increments immediately as XML data arrives.
"""
self._buffer += new_text
current_text = self._buffer
# Check if we have a tool call
has_tool_call = self.bot_token in current_text
if not has_tool_call:
# Check if buffer could be the start of a tool call
# Keep buffer if it could be a partial match of bot_token
is_potential_start = any(
self.bot_token.startswith(current_text[-i:])
for i in range(1, min(len(current_text), len(self.bot_token)) + 1)
)
if not is_potential_start:
# Not a potential tool call, return as normal text
# Must return the entire buffer (current_text), not just new_text,
# because buffer may contain previously accumulated characters like '<'
# that turned out not to be part of a tool call
output_text = current_text
self._buffer = ""
if self.eot_token in output_text:
output_text = output_text.replace(self.eot_token, "")
return StreamingParseResult(normal_text=output_text)
else:
# Could be start of tool call, keep buffering
return StreamingParseResult(normal_text="", calls=[])
if not hasattr(self, "_tool_indices"):
self._tool_indices = self._get_tool_indices(tools)
calls: list[ToolCallItem] = []
try:
# Try to match a partial or complete tool call
partial_match = re.search(
pattern=r"<tool_call>(.*?)(<arg_key>.*?)?(</tool_call>|$)",
string=current_text,
flags=re.DOTALL,
)
if partial_match:
func_name = partial_match.group(1).strip()
func_args_raw = partial_match.group(2).strip()
is_tool_end = partial_match.group(3)
# 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 = [""]
self._streamed_raw_length = 0
self.current_tool_name_sent = False
self._reset_streaming_state()
# 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("")
# Send tool name first if not sent yet
if not self.current_tool_name_sent:
assert func_name, "func_name should not be empty"
calls.append(
ToolCallItem(
tool_index=self.current_tool_id,
name=func_name,
parameters="",
)
)
self.current_tool_name_sent = True
self._streamed_raw_length = 0
self._reset_streaming_state()
# Store the tool call info
self.prev_tool_call_arr[self.current_tool_id] = {
"name": func_name,
"arguments": {},
}
else:
# Process XML to JSON streaming
current_raw_length = len(func_args_raw)
if current_raw_length > self._streamed_raw_length:
# Get the new raw XML content
raw_increment = func_args_raw[self._streamed_raw_length :]
# Convert XML increment to JSON increment using state machine
json_increment = self._process_xml_to_json_streaming(
raw_increment, func_name, tools
)
if json_increment:
calls.append(
ToolCallItem(
tool_index=self.current_tool_id,
name=None,
parameters=json_increment,
)
)
self._last_arguments += json_increment
self.streamed_args_for_tool[
self.current_tool_id
] += json_increment
# Update the streamed length
self._streamed_raw_length = current_raw_length
if is_tool_end == self.eot_token:
if self._is_first_param:
empty_object = "{}"
calls.append(
ToolCallItem(
tool_index=self.current_tool_id,
name=None,
parameters=empty_object,
)
)
self._last_arguments += empty_object
elif not self._last_arguments.endswith("}"):
closing_brace = "}"
calls.append(
ToolCallItem(
tool_index=self.current_tool_id,
name=None,
parameters=closing_brace,
)
)
self._last_arguments += closing_brace
self.streamed_args_for_tool[
self.current_tool_id
] += closing_brace
try:
pairs = self.func_arg_regex.findall(func_args_raw)
if pairs:
arguments = self._parse_argument_pairs(
pairs, func_name, tools
)
self.prev_tool_call_arr[self.current_tool_id][
"arguments"
] = arguments
except Exception as e:
logger.debug(
f"Failed to parse arguments: {e}", exc_info=True
)
# Remove the completed tool call from buffer
self._buffer = current_text[partial_match.end(3) :]
result = StreamingParseResult(normal_text="", calls=calls)
self.current_tool_id += 1
self._last_arguments = ""
self.current_tool_name_sent = False
self._streamed_raw_length = 0
self._reset_streaming_state()
return result
return StreamingParseResult(normal_text="", calls=calls)
except Exception as e:
logger.error(f"Error in parse_streaming_increment: {e}", exc_info=True)
return StreamingParseResult(normal_text=current_text)
def _parse_argument_pairs(
self, pairs: List[Tuple[str, str]], func_name: str, tools: List[Tool]
) -> Dict[str, Any]:
"""Parse argument key-value pairs with type coercion.
Args:
pairs: List of (key, value) tuples from regex matching
func_name: Name of the function
tools: List of available tools
Returns:
Dictionary of parsed arguments
"""
arguments = {}
for arg_key, arg_value in pairs:
arg_key = arg_key.strip()
arg_value = arg_value.strip()
arg_type = get_argument_type(func_name, arg_key, tools)
parsed_value, is_good_json = parse_arguments(arg_value, arg_type)
if arg_type == "string":
# Only convert to string if explicitly defined as string type
if isinstance(parsed_value, str):
arguments[arg_key] = parsed_value
elif isinstance(parsed_value, (dict, list)):
# If parsed as dict/list but schema says string, convert to JSON string
arguments[arg_key] = json.dumps(parsed_value, ensure_ascii=False)
else:
arguments[arg_key] = str(parsed_value)
elif arg_type is None:
# If type is not defined, keep the parsed value as-is
arguments[arg_key] = parsed_value if is_good_json else arg_value
else:
# For other types (number, object, array, etc.), use parsed value
arguments[arg_key] = parsed_value if is_good_json else arg_value
return arguments
def supports_structural_tag(self) -> bool:
return False
def structure_info(self) -> _GetInfoFunc:
raise NotImplementedError()

View File

@@ -43,9 +43,12 @@ def get_argument_type(
if func_name not in name2tool:
return None
tool = name2tool[func_name]
if arg_key not in tool.function.parameters["properties"]:
properties = (tool.function.parameters or {}).get("properties", {})
if not isinstance(properties, dict):
properties = {}
if arg_key not in properties:
return None
return tool.function.parameters["properties"][arg_key].get("type", None)
return properties[arg_key].get("type", None)
def _convert_to_number(value: str) -> Any:

View File

@@ -12,7 +12,7 @@
# limitations under the License.
# ==============================================================================
"""Inference-only GLM-4.5, GLM-4.6 model compatible with HuggingFace weights"""
"""Inference-only GLM-4.5, GLM-4.6 and GLM-4.7 model compatible with HuggingFace weights"""
import logging
from typing import Any, Dict, Iterable, List, Optional, Tuple, Union